{"audited_on":"2026-11-06","baseline":"1.16.0","version":"2.2","revision":"232484765b25c3f8f1eee656cb8c728616ebac6a","method":"Source and official documentation audit of the recorded revision, with 1.16.0 kept as the baseline. Selected integrations include live HTTP or WebSocket recordings and focused regressions; each cell separates implementation support from validation evidence. The comparison covers application-owned conversations, tools, media, embeddings, transcription, speech and batches. Realtime conversations and provider-managed conversation storage are outside the scope; their provider offerings remain recorded but are excluded from comparison totals. One-shot streaming speech and transcription remain in scope. Access restrictions and upstream errors are recorded explicitly. Provider offerings are not an exhaustive endpoint inventory.","groups":[{"id":"conversation","name":"Conversations","summary":true},{"id":"tools","name":"Tools","summary":true},{"id":"context","name":"Caching and Context","summary":true},{"id":"media","name":"Media","summary":true},{"id":"search","name":"Documents and Search","summary":true},{"id":"resources","name":"Files and Batches","summary":true},{"id":"additional","name":"Additional Provider APIs","summary":false},{"id":"boundaries","name":"Realtime and Training","summary":false}],"statuses":{"native":{"symbol":"✓","label":"Built in","description":"A RubyLLM API handles the feature on supported models and protocols. Selecting a protocol or passing options to a registered alias does not make the integration partial. The cell records model and endpoint restrictions."},"partial":{"symbol":"◐","label":"Partial","description":"Some paths or result formats are supported; others are missing or need operational verification. The cell explains the limitation."},"passthrough":{"symbol":"{}","label":"Raw options","description":"Requires provider-specific request options, raw content, or a tool definition rather than a dedicated RubyLLM interface."},"missing":{"symbol":"×","label":"Missing","description":"The provider offers this feature, but the audited version has no integration for it."},"na":{"symbol":"–","label":"Not in this API","description":"The audited API inventory does not document this operation, or a live check found the documented operation is not served. The cell identifies the scope; separate products, client-side tools and chat prompts doing a similar task are not counted."},"unknown":{"symbol":"?","label":"Unverified","description":"Provider availability or implementation behavior remains unresolved. The cell records the specific uncertainty and the sources checked."},"deprecated":{"symbol":"–","label":"Deprecated","description":"The provider has deprecated this API and scheduled its removal. It is excluded from missing-integration work; the cell records the shutdown date."},"out_of_scope":{"symbol":"╱","label":"Outside scope","description":"The provider offers this feature through realtime conversations or provider-managed conversation state. These integrations are outside RubyLLM's coverage and excluded from the comparison totals."}},"features":[{"id":"chat","name":"Chat","group":"conversation"},{"id":"streaming","name":"Chat streaming","group":"conversation"},{"id":"structured_output","name":"Structured output","group":"conversation"},{"id":"vision","name":"Image input","group":"conversation"},{"id":"pdf_input","name":"PDF input","group":"conversation"},{"id":"audio_input","name":"Audio input","group":"conversation"},{"id":"video_input","name":"Video input","group":"conversation"},{"id":"thinking","name":"Thinking results","group":"conversation"},{"id":"thinking_controls","name":"Thinking controls","group":"conversation"},{"id":"citations","name":"Citations","group":"conversation"},{"id":"tools","name":"Function tools","group":"tools"},{"id":"tool_choice","name":"Tool choice","group":"tools"},{"id":"parallel_tools","name":"Multiple tool calls","group":"tools"},{"id":"parallel_tool_control","name":"Limit parallel calls","group":"tools"},{"id":"server_web_search","name":"Web search tool","group":"tools"},{"id":"server_web_fetch","name":"Web fetch tool","group":"tools"},{"id":"server_code_execution","name":"Code execution tool","group":"tools"},{"id":"server_file_search","name":"File search tool","group":"tools"},{"id":"server_mcp","name":"Remote MCP tools","group":"tools"},{"id":"prompt_caching","name":"Prompt caching","group":"context"},{"id":"cache_boundaries","name":"Cache boundaries","group":"context"},{"id":"explicit_cache","name":"Managed cache resources","group":"context"},{"id":"compaction","name":"Context compaction","group":"context"},{"id":"token_counting","name":"Token counting","group":"context"},{"id":"image_generation","name":"Image generation","group":"media"},{"id":"image_editing","name":"Image editing","group":"media"},{"id":"video_generation","name":"Video generation","group":"media"},{"id":"speech","name":"Speech generation","group":"media"},{"id":"transcription","name":"Transcription","group":"media"},{"id":"streaming_transcription","name":"Streaming transcription","group":"media"},{"id":"diarization","name":"Speaker identification","group":"media"},{"id":"moderation","name":"Moderation","group":"media"},{"id":"ocr","name":"OCR / document parsing","group":"search"},{"id":"embeddings","name":"Text embeddings","group":"search"},{"id":"multimodal_embeddings","name":"Multimodal embeddings","group":"search"},{"id":"reranking","name":"Reranking","group":"search"},{"id":"file_upload","name":"File upload","group":"resources"},{"id":"file_download","name":"File download","group":"resources"},{"id":"chat_batches","name":"Chat batches","group":"resources"},{"id":"embedding_batches","name":"Embedding batches","group":"resources"},{"id":"advisor_tool","name":"Advisor tool","group":"additional"},{"id":"agent_responses","name":"Perplexity Agent API","group":"additional"},{"id":"agent_skills","name":"Agent skills","group":"additional"},{"id":"async_research","name":"Asynchronous research","group":"additional"},{"id":"background_responses","name":"Background Responses jobs","group":"additional"},{"id":"browser_use","name":"Browser use","group":"additional"},{"id":"classification","name":"Classification","group":"additional"},{"id":"computer_use","name":"Computer use","group":"additional"},{"id":"context_editing","name":"Context editing","group":"additional"},{"id":"contextual_embeddings","name":"Contextual embeddings","group":"additional"},{"id":"dedicated_transcription_api","name":"Dedicated transcription API","group":"additional"},{"id":"dubbing","name":"Dubbing","group":"additional"},{"id":"fast_inference","name":"Fast inference","group":"additional"},{"id":"file_management","name":"File listing and deletion","group":"additional"},{"id":"file_search_store_management","name":"File search store management","group":"additional"},{"id":"fill_in_middle","name":"Fill-in-the-middle completion","group":"additional"},{"id":"fine_tuning","name":"Fine-tuning","group":"boundaries"},{"id":"google_maps","name":"Google Maps grounding","group":"additional"},{"id":"guardrail_configuration","name":"Guardrail configuration","group":"additional"},{"id":"hosted_conversations","name":"Hosted conversations","group":"additional"},{"id":"image_batches","name":"Image batches","group":"additional"},{"id":"interactions_api","name":"Interactions API","group":"additional"},{"id":"json_mode","name":"JSON object mode","group":"additional"},{"id":"mantle_protocols","name":"Bedrock Mantle protocols","group":"additional"},{"id":"manual_compaction","name":"Manual context compaction","group":"additional"},{"id":"music_generation","name":"Music generation","group":"additional"},{"id":"nonchat_batches","name":"Other batch operations","group":"additional"},{"id":"omni_video_generation","name":"Omni video generation","group":"additional"},{"id":"partner_model_protocols","name":"Partner model protocols","group":"additional"},{"id":"prefix_completion","name":"Prefix completion","group":"additional"},{"id":"programmatic_tool_calling","name":"Programmatic tool calling","group":"additional"},{"id":"prompt_cache_controls","name":"Prompt cache options","group":"additional"},{"id":"provider_server_fallback","name":"Provider-managed fallback","group":"additional"},{"id":"realtime","name":"Realtime sessions","group":"boundaries"},{"id":"response_caching","name":"Response caching","group":"additional"},{"id":"responses","name":"Responses API","group":"additional"},{"id":"responses_batches","name":"Responses batches","group":"additional"},{"id":"search_api","name":"Standalone search API","group":"additional"},{"id":"server_apply_patch","name":"Apply patch tool","group":"additional"},{"id":"server_tool_image_generation","name":"Image generation tool","group":"additional"},{"id":"server_tool_search","name":"Tool search","group":"additional"},{"id":"server_x_search","name":"X search tool","group":"additional"},{"id":"shell_computer_tools","name":"Shell and computer tools","group":"additional"},{"id":"sound_effects","name":"Sound effects","group":"additional"},{"id":"streaming_speech","name":"Streaming speech generation","group":"additional"},{"id":"task_budgets","name":"Task budgets","group":"additional"},{"id":"text_intelligence","name":"Text analysis API","group":"additional"},{"id":"vector_store_management","name":"Vector store management","group":"additional"},{"id":"video_batches","name":"Video batches","group":"additional"},{"id":"video_editing","name":"Video editing","group":"additional"},{"id":"video_extension","name":"Video extension","group":"additional"},{"id":"voice_cloning","name":"Voice cloning","group":"additional"},{"id":"voice_conversion","name":"Voice conversion","group":"additional"},{"id":"web_fetch_api","name":"Standalone web fetch API","group":"additional"},{"id":"websockets","name":"Responses over WebSocket","group":"boundaries"}],"summary":{"feature_count":40,"groups":[{"id":"conversation","name":"Conversations","v1":91,"v2":169,"offered":169},{"id":"tools","name":"Tools","v1":44,"v2":125,"offered":128},{"id":"context","name":"Caching and Context","v1":12,"v2":51,"offered":54},{"id":"media","name":"Media","v1":15,"v2":91,"offered":91},{"id":"search","name":"Documents and Search","v1":8,"v2":31,"offered":31},{"id":"resources","name":"Files and Batches","v1":0,"v2":51,"offered":51}],"maximum":169},"sources":{"s3":{"title":"OpenAI using tools","url":"https://developers.openai.com/api/docs/guides/tools"},"s4":{"title":"OpenAI prompt caching","url":"https://developers.openai.com/api/docs/guides/prompt-caching"},"s6":{"title":"OpenAI video generation","url":"https://developers.openai.com/api/docs/guides/video-generation"},"s7":{"title":"OpenAI Batch API","url":"https://developers.openai.com/api/docs/guides/batch"},"s8":{"title":"Azure OpenAI Responses","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses"},"s10":{"title":"Azure Responses web search","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/web-search"},"s11":{"title":"Azure prompt caching","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/prompt-caching"},"s14":{"title":"Azure Sora video generation","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/video-generation"},"s15":{"title":"Azure global batch processing","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch"},"s16":{"title":"Azure batch REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/batch"},"s17":{"title":"xAI tools overview","url":"https://docs.x.ai/developers/tools/overview"},"s18":{"title":"xAI models","url":"https://docs.x.ai/developers/models"},"s19":{"title":"xAI files","url":"https://docs.x.ai/developers/files"},"s20":{"title":"xAI Voice reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/voice"},"s21":{"title":"xAI prompt caching","url":"https://docs.x.ai/developers/advanced-api-usage/prompt-caching"},"s22":{"title":"xAI context compaction","url":"https://docs.x.ai/developers/advanced-api-usage/context-compaction"},"s23":{"title":"xAI account and tokenize reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/other"},"s26":{"title":"xAI Batch API","url":"https://docs.x.ai/developers/advanced-api-usage/batch-api"},"s27":{"title":"xAI release notes","url":"https://docs.x.ai/developers/release-notes"},"s33":{"title":"Claude features overview","url":"https://platform.claude.com/docs/en/build-with-claude/overview"},"s34":{"title":"Embeddings","url":"https://platform.claude.com/docs/en/build-with-claude/embeddings"},"s35":{"title":"Claude citations","url":"https://platform.claude.com/docs/en/build-with-claude/citations"},"s36":{"title":"MCP connector","url":"https://platform.claude.com/docs/en/agents-and-tools/mcp-connector"},"s37":{"title":"Claude prompt caching","url":"https://platform.claude.com/docs/en/build-with-claude/prompt-caching"},"s38":{"title":"Compaction overview","url":"https://platform.claude.com/docs/en/build-with-claude/compaction"},"s39":{"title":"Files API","url":"https://platform.claude.com/docs/en/build-with-claude/files"},"s40":{"title":"Claude tool search tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool"},"s70":{"title":"Converse API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Converse.html"},"s72":{"title":"Bedrock prompt caching","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html"},"s73":{"title":"CountTokens API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_CountTokens.html"},"s74":{"title":"Titan Text Embeddings","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/titan-embedding-models.html"},"s76":{"title":"Rerank API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agent-runtime_Rerank.html"},"s79":{"title":"Nova Sonic bidirectional streaming","url":"https://docs.aws.amazon.com/nova/latest/userguide/speech-bidirection.html"},"s81":{"title":"Use the ApplyGuardrail API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-use-independent-api.html"},"s82":{"title":"Bedrock batch inference","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference.html"},"s83":{"title":"Bedrock model customization","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html"},"s84":{"title":"Cohere Chat API","url":"https://docs.cohere.com/reference/chat"},"s85":{"title":"Cohere Parse API","url":"https://docs.cohere.com/reference/parse"},"s86":{"title":"Cohere tokenize","url":"https://docs.cohere.com/reference/tokenize"},"s87":{"title":"Cohere reasoning","url":"https://docs.cohere.com/docs/reasoning"},"s88":{"title":"Cohere Embed models","url":"https://docs.cohere.com/docs/cohere-embed"},"s89":{"title":"Cohere Rerank API","url":"https://docs.cohere.com/reference/rerank"},"s90":{"title":"Cohere create transcription","url":"https://docs.cohere.com/reference/create-audio-transcription"},"s91":{"title":"Cohere Embed Jobs guide","url":"https://docs.cohere.com/docs/embed-jobs-api"},"s92":{"title":"Mistral API reference","url":"https://docs.mistral.ai/api"},"s93":{"title":"Mistral Speech API","url":"https://docs.mistral.ai/api/endpoint/audio/speech"},"s94":{"title":"Mistral Transcriptions API","url":"https://docs.mistral.ai/api/endpoint/audio/transcriptions"},"s95":{"title":"Mistral OCR API","url":"https://docs.mistral.ai/api/endpoint/ocr"},"s96":{"title":"Mistral Files API","url":"https://docs.mistral.ai/api/endpoint/files"},"s99":{"title":"Mistral deprecated fine-tuning","url":"https://docs.mistral.ai/resources/deprecated/finetuning"},"s100":{"title":"Pre-recorded audio reference","url":"https://developers.deepgram.com/reference/speech-to-text/listen-pre-recorded"},"s101":{"title":"Deepgram text to speech getting started","url":"https://developers.deepgram.com/docs/text-to-speech"},"s102":{"title":"Diarization","url":"https://developers.deepgram.com/docs/diarization"},"s103":{"title":"Live audio reference","url":"https://developers.deepgram.com/reference/speech-to-text/listen-streaming"},"s104":{"title":"Twilio and Deepgram Voice Agent","url":"https://developers.deepgram.com/docs/twilio-and-deepgram-voice-agent"},"s105":{"title":"Text Intelligence intent recognition","url":"https://developers.deepgram.com/docs/text-intention-recognition"},"s107":{"title":"Create transcript","url":"https://elevenlabs.io/docs/api-reference/speech-to-text/convert"},"s108":{"title":"Realtime speech to text","url":"https://elevenlabs.io/docs/api-reference/speech-to-text/v-1-speech-to-text-realtime"},"s109":{"title":"Generate audio in real-time","url":"https://elevenlabs.io/docs/eleven-api/guides/how-to/websockets/realtime-tts"},"s110":{"title":"Compose music","url":"https://elevenlabs.io/docs/api-reference/music/compose"},"s111":{"title":"Voice changer","url":"https://elevenlabs.io/docs/api-reference/speech-to-speech/convert"},"s115":{"title":"Perplexity Embeddings API","url":"https://docs.perplexity.ai/docs/embeddings/quickstart"},"s118":{"title":"Audio","url":"https://openrouter.ai/docs/guides/overview/multimodal/audio"},"s119":{"title":"Video Generation","url":"https://openrouter.ai/docs/guides/overview/multimodal/video-generation"},"s122":{"title":"Prompt Caching","url":"https://openrouter.ai/docs/guides/best-practices/prompt-caching"},"s123":{"title":"Message Transforms","url":"https://openrouter.ai/docs/guides/features/message-transforms"},"s126":{"title":"Image Generation","url":"https://openrouter.ai/docs/guides/overview/multimodal/image-generation"},"s127":{"title":"Text-to-Speech","url":"https://openrouter.ai/docs/guides/overview/multimodal/tts"},"s128":{"title":"Speech-to-Text","url":"https://openrouter.ai/docs/guides/overview/multimodal/stt"},"s129":{"title":"Batch API quickstart","url":"https://openrouter.ai/docs/batch-quickstart"},"s130":{"title":"Response Caching","url":"https://openrouter.ai/docs/guides/features/response-caching"},"s131":{"title":"Ollama OpenAI compatibility","url":"https://docs.ollama.com/api/openai-compatibility"},"s132":{"title":"Ollama structured outputs","url":"https://docs.ollama.com/capabilities/structured-outputs"},"s133":{"title":"Ollama web search and fetch APIs","url":"https://docs.ollama.com/capabilities/web-search"},"s135":{"title":"Ollama Cloud","url":"https://docs.ollama.com/cloud"},"s136":{"title":"GPUStack inference APIs","url":"https://docs.gpustack.ai/latest/integrations/inference-apis/"},"s137":{"title":"OpenAI API docs index","url":"https://developers.openai.com/api/docs/llms.txt"},"s138":{"title":"OpenAI skills","url":"https://developers.openai.com/api/docs/guides/tools-skills"},"s140":{"title":"OpenAI background mode","url":"https://developers.openai.com/api/docs/guides/background"},"s141":{"title":"OpenAI computer use","url":"https://developers.openai.com/api/docs/guides/tools-computer-use"},"s142":{"title":"OpenAI apply patch","url":"https://developers.openai.com/api/docs/guides/tools-apply-patch"},"s143":{"title":"OpenAI file transcription","url":"https://developers.openai.com/api/docs/guides/speech-to-text"},"s144":{"title":"OpenAI Fast mode","url":"https://developers.openai.com/api/docs/guides/fast-mode"},"s145":{"title":"OpenAI file search","url":"https://developers.openai.com/api/docs/guides/tools-file-search"},"s146":{"title":"OpenAI tool search","url":"https://developers.openai.com/api/docs/guides/tools-tool-search"},"s147":{"title":"OpenAI text to speech","url":"https://developers.openai.com/api/docs/guides/text-to-speech"},"s148":{"title":"OpenAI web search","url":"https://developers.openai.com/api/docs/guides/tools-web-search"},"s149":{"title":"Azure OpenAI image, audio and video REST reference (v1 preview)","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/reference-preview-latest"},"s151":{"title":"Cohere reranking on Azure AI Foundry","url":"https://docs.cohere.com/v2/docs/cohere-on-azure/azure-ai-reranking"},"s152":{"title":"Cohere models on Azure","url":"https://docs.cohere.com/v2/docs/cohere-on-microsoft-azure"},"s153":{"title":"Azure Responses WebSocket mode","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/websockets"},"s155":{"title":"Azure AI Content Safety","url":"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/"},"s156":{"title":"Azure Document Intelligence overview","url":"https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/overview"},"s157":{"title":"xAI docs index","url":"https://docs.x.ai/llms.txt"},"s159":{"title":"xAI deferred chat completions","url":"https://docs.x.ai/developers/advanced-api-usage/deferred-chat-completions"},"s160":{"title":"xAI speech to text","url":"https://docs.x.ai/developers/model-capabilities/audio/speech-to-text"},"s161":{"title":"xAI priority processing","url":"https://docs.x.ai/developers/advanced-api-usage/priority-processing"},"s162":{"title":"xAI Collections API guide","url":"https://docs.x.ai/developers/files/collections/api"},"s163":{"title":"xAI custom voices","url":"https://docs.x.ai/developers/model-capabilities/audio/custom-voices"},"s164":{"title":"DeepSeek API docs sitemap","url":"https://api-docs.deepseek.com/sitemap.xml"},"s165":{"title":"DeepSeek Responses API guide","url":"https://api-docs.deepseek.com/guides/responses_api"},"s166":{"title":"DeepSeek Responses reference","url":"https://api-docs.deepseek.com/api/create-response"},"s167":{"title":"DeepSeek Token and Token Usage","url":"https://api-docs.deepseek.com/quick_start/token_usage"},"s168":{"title":"DeepSeek Files API guide","url":"https://api-docs.deepseek.com/guides/files_api"},"s169":{"title":"Cohere documentation sitemap","url":"https://docs.cohere.com/sitemap.xml"},"s170":{"title":"Cohere deprecations","url":"https://docs.cohere.com/docs/deprecations"},"s171":{"title":"Cohere list datasets","url":"https://docs.cohere.com/reference/list-datasets"},"s172":{"title":"Cohere delete dataset","url":"https://docs.cohere.com/reference/delete-dataset"},"s173":{"title":"Cohere create batch","url":"https://docs.cohere.com/reference/create-batch"},"s174":{"title":"Cohere get batch","url":"https://docs.cohere.com/reference/get-batch"},"s176":{"title":"Mistral Skills API","url":"https://docs.mistral.ai/api/endpoint/beta/skills"},"s177":{"title":"Mistral Libraries API","url":"https://docs.mistral.ai/api/endpoint/beta/libraries"},"s178":{"title":"Mistral RAG search-index API","url":"https://docs.mistral.ai/api/endpoint/beta/rag/search_indexes"},"s179":{"title":"Mistral Conversations API","url":"https://docs.mistral.ai/api/endpoint/beta/conversations"},"s180":{"title":"Mistral image generation tool","url":"https://docs.mistral.ai/studio/agents/agent-tools/image_generation"},"s181":{"title":"Mistral Voices API","url":"https://docs.mistral.ai/api/endpoint/audio/voices"},"s182":{"title":"Mistral web search tool","url":"https://docs.mistral.ai/studio/agents/agent-tools/websearch"},"s183":{"title":"Mistral Connectors API","url":"https://docs.mistral.ai/api/endpoint/beta/connectors"},"s185":{"title":"Mistral Embeddings API","url":"https://docs.mistral.ai/api/endpoint/embeddings"},"s186":{"title":"Claude API overview","url":"https://platform.claude.com/docs/en/api/overview"},"s187":{"title":"Structured outputs","url":"https://platform.claude.com/docs/en/build-with-claude/structured-outputs"},"s188":{"title":"Claude API primer","url":"https://platform.claude.com/docs/en/claude_api_primer"},"s202":{"title":"generateContent request and tool schemas","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/mcp/tools_list/generate_content"},"s203":{"title":"Interactions API reference","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/interactions-api"},"s204":{"title":"Interact with managed agents","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/managed-agents/interact-with-agents"},"s205":{"title":"Deep Research agent","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/agents/use-deep-research"},"s206":{"title":"Grok Responses API","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/grok/responses"},"s207":{"title":"Priority PayGo","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/priority-paygo"},"s208":{"title":"Claude fast mode","url":"https://platform.claude.com/docs/en/build-with-claude/fast-mode"},"s210":{"title":"RAG Engine API","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/rag-api-v1"},"s212":{"title":"Extend videos","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/extend-videos"},"s213":{"title":"Edit videos","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/edit-videos"},"s214":{"title":"Context caching overview","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/context-cache/context-cache-overview"},"s215":{"title":"Gemini Live API","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/live-api"},"s217":{"title":"Claude task budgets","url":"https://platform.claude.com/docs/en/build-with-claude/task-budgets"},"s220":{"title":"Amazon Bedrock API operations","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_Operations.html"},"s222":{"title":"Bedrock API compatibility","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/models-api-compatibility.html"},"s223":{"title":"Bedrock latency-optimized inference","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/latency-optimized-inference.html"},"s224":{"title":"S3 API operations","url":"https://docs.aws.amazon.com/AmazonS3/latest/API/API_Operations_Amazon_Simple_Storage_Service.html"},"s225":{"title":"CreateKnowledgeBase API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agent_CreateKnowledgeBase.html"},"s226":{"title":"Claude Messages API on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages.html"},"s227":{"title":"Bedrock Web Search","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/web-search.html"},"s229":{"title":"Bedrock server-side tool use","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/tool-use-server-side.html"},"s230":{"title":"Voxtral Small 24B model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-mistral-ai-voxtral-small-24b-2507.html"},"s231":{"title":"Configure the Voice Agent","url":"https://developers.deepgram.com/docs/configure-voice-agent"},"s232":{"title":"Voice Agent feature overview","url":"https://developers.deepgram.com/docs/voice-agent-feature-overview"},"s233":{"title":"Deepgram WebSocket API specification","url":"https://developers.deepgram.com/asyncapi.json"},"s234":{"title":"Deepgram documentation index","url":"https://developers.deepgram.com/llms.txt"},"s235":{"title":"Deepgram REST API specification","url":"https://developers.deepgram.com/openapi.json"},"s236":{"title":"Deepgram API overview","url":"https://developers.deepgram.com/reference/deepgram-api-overview"},"s237":{"title":"Voice Agent conversation context","url":"https://developers.deepgram.com/docs/voice-agent-conversation-context"},"s238":{"title":"Channel behavior","url":"https://elevenlabs.io/docs/eleven-agents/customization/channel-behavior"},"s239":{"title":"Tools","url":"https://elevenlabs.io/docs/eleven-agents/customization/tools"},"s240":{"title":"Client events","url":"https://elevenlabs.io/docs/eleven-agents/customization/events/client-events"},"s241":{"title":"ElevenLabs complete documentation and API inventory","url":"https://elevenlabs.io/docs/llms.txt"},"s242":{"title":"Multimodal input","url":"https://elevenlabs.io/docs/eleven-agents/customization/multimodal-input"},"s243":{"title":"Models","url":"https://elevenlabs.io/docs/eleven-agents/customization/llm"},"s244":{"title":"Create agent","url":"https://elevenlabs.io/docs/eleven-agents/api-reference/agents/create"},"s245":{"title":"Get conversation details","url":"https://elevenlabs.io/docs/eleven-agents/api-reference/conversations/get"},"s246":{"title":"Retrieval-Augmented Generation","url":"https://elevenlabs.io/docs/eleven-agents/customization/knowledge-base/rag"},"s247":{"title":"Parallel","url":"https://elevenlabs.io/docs/eleven-agents/customization/integrations/parallel"},"s248":{"title":"Code tools","url":"https://elevenlabs.io/docs/eleven-agents/customization/tools/code-tools"},"s249":{"title":"ElevenAgents MCP tools","url":"https://elevenlabs.io/docs/eleven-agents/customization/tools/mcp"},"s250":{"title":"Create Image Generation","url":"https://elevenlabs.io/docs/api-reference/flows/image/create"},"s251":{"title":"Create Video Generation","url":"https://elevenlabs.io/docs/api-reference/flows/video/create"},"s252":{"title":"Procedures","url":"https://elevenlabs.io/docs/eleven-agents/customization/procedures"},"s253":{"title":"ElevenAgents conversation analysis","url":"https://elevenlabs.io/docs/eleven-agents/customization/agent-analysis"},"s254":{"title":"Create Music Finetune","url":"https://elevenlabs.io/docs/api-reference/music/finetunes/create"},"s255":{"title":"Guardrails","url":"https://elevenlabs.io/docs/eleven-agents/best-practices/guardrails"},"s256":{"title":"Compute RAG index in batch","url":"https://elevenlabs.io/docs/eleven-agents/api-reference/knowledge-base/compute-rag-index-batch"},"s257":{"title":"Image & Video","url":"https://elevenlabs.io/docs/overview/capabilities/image-video"},"s258":{"title":"Perplexity Orchestrator API (OpenAI-compatible) specification","url":"https://docs.perplexity.ai/openapi-gateway-chat.json"},"s260":{"title":"Perplexity Orchestrator API (OpenAI Responses-compatible) specification","url":"https://docs.perplexity.ai/openapi-gateway-responses.json"},"s261":{"title":"Perplexity Orchestrator API","url":"https://docs.perplexity.ai/docs/router/quickstart"},"s262":{"title":"Agent API tools overview","url":"https://docs.perplexity.ai/docs/agent-api/tools/overview"},"s263":{"title":"Create Agent Response","url":"https://docs.perplexity.ai/api-reference/agent-post"},"s264":{"title":"Perplexity documentation index","url":"https://docs.perplexity.ai/llms.txt"},"s265":{"title":"Perplexity API specification","url":"https://docs.perplexity.ai/openapi.json"},"s266":{"title":"Perplexity Orchestrator API (Anthropic-compatible) specification","url":"https://docs.perplexity.ai/openapi-gateway-messages.json"},"s267":{"title":"Agent API conversation state","url":"https://docs.perplexity.ai/docs/agent-api/conversation-state"},"s268":{"title":"Orchestrator routing and reliability","url":"https://docs.perplexity.ai/docs/router/routing-and-reliability"},"s269":{"title":"Server Tools","url":"https://openrouter.ai/docs/guides/features/server-tools"},"s270":{"title":"Shell server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/shell"},"s271":{"title":"Bash server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/bash"},"s272":{"title":"Tool Search server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/tool-search"},"s273":{"title":"Create a response","url":"https://openrouter.ai/docs/api/api-reference/responses/create-a-response"},"s274":{"title":"OpenRouter documentation index","url":"https://openrouter.ai/docs/llms.txt"},"s275":{"title":"PDF Inputs","url":"https://openrouter.ai/docs/guides/overview/multimodal/pdfs"},"s276":{"title":"Responses API overview","url":"https://openrouter.ai/docs/api_reference/responses/overview"},"s277":{"title":"Custom Classifiers","url":"https://openrouter.ai/docs/guides/features/classifiers"},"s278":{"title":"Service Tiers","url":"https://openrouter.ai/docs/guides/features/service-tiers"},"s279":{"title":"Files API","url":"https://openrouter.ai/docs/guides/features/files-api"},"s281":{"title":"Submit a video generation request","url":"https://openrouter.ai/docs/api/api-reference/video-generation/submit-a-video-generation-request"},"s282":{"title":"Model Fallbacks","url":"https://openrouter.ai/docs/guides/routing/model-fallbacks"},"s283":{"title":"Ollama documentation index","url":"https://docs.ollama.com/llms.txt"},"s286":{"title":"Ollama generate API","url":"https://docs.ollama.com/api/generate"},"s291":{"title":"Ollama Cloud model catalog","url":"https://ollama.com/api/tags"},"s292":{"title":"GPUStack built-in inference backends","url":"https://docs.gpustack.ai/latest/user-guide/built-in-inference-backends/"},"s293":{"title":"vLLM Responses client with MCP tools","url":"https://docs.vllm.ai/en/latest/examples/tool_calling/openai_responses_client_with_mcp_tools/"},"s294":{"title":"GPUStack LMCache evaluation","url":"https://docs.gpustack.ai/latest/performance-lab/references/evaluating-lmcache-prefill-acceleration-in-vllm/"},"s295":{"title":"vLLM OpenAI-compatible server","url":"https://docs.vllm.ai/en/latest/serving/online_serving/openai_compatible_server/"},"s296":{"title":"GPUStack audio models","url":"https://docs.gpustack.ai/latest/using-models/using-audio-models/"},"s297":{"title":"vLLM speech to text APIs","url":"https://docs.vllm.ai/en/latest/serving/online_serving/speech_to_text/"},"s300":{"title":"vLLM embedding usages","url":"https://docs.vllm.ai/en/latest/models/pooling_models/embed/"},"s301":{"title":"Mistral agent tools","url":"https://docs.mistral.ai/studio/agents/agent-tools"},"s302":{"title":"Mistral connectors in completions cookbook","url":"https://docs.mistral.ai/resources/cookbooks/mistral-connectors-05-connectors-in-completions"},"s304":{"title":"Embeddings","url":"https://openrouter.ai/docs/api_reference/embeddings"},"s306":{"title":"OpenAI counting tokens","url":"https://developers.openai.com/api/docs/guides/token-counting"},"s307":{"title":"OpenAI input token count reference","url":"https://developers.openai.com/api/reference/resources/responses/subresources/input_tokens/methods/count"},"s308":{"title":"Generate videos from text","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-text"},"s309":{"title":"Generate videos from an image","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-an-image"},"s312":{"title":"Azure OpenAI image and audio REST reference (2025-04-01-preview)","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/reference-preview"},"s313":{"title":"Image generation models on Azure","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/dall-e"},"s314":{"title":"Orchestrator models and pricing","url":"https://docs.perplexity.ai/docs/router/models"},"s318":{"title":"Get multimodal embeddings","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/get-multimodal-embeddings"},"s325":{"title":"Cohere Parse quickstart","url":"https://docs.cohere.com/v2/docs/parse-quickstart"},"s326":{"title":"Cohere models","url":"https://docs.cohere.com/docs/models"},"s327":{"title":"Cohere Parse release notes","url":"https://docs.cohere.com/changelog/parse"},"s329":{"title":"vLLM multimodal inputs","url":"https://docs.vllm.ai/en/stable/features/multimodal_inputs/"},"s331":{"title":"Stable Diffusion 3.5 Large parameters","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-diffusion-3-5-large.html"},"s332":{"title":"Stable Image Core parameters","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-diffusion-stable-image-core-text-image-request-response.html"},"s333":{"title":"InvokeModel API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html"},"s334":{"title":"Stability AI Image Services","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/stable-image-services.html"},"s335":{"title":"Stable Image Inpaint model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-stability-ai-stable-image-inpaint.html"},"s337":{"title":"OpenAI MCP servers","url":"https://developers.openai.com/api/docs/guides/tools-connectors-mcp"},"s338":{"title":"Microsoft Learn MCP server","url":"https://learn.microsoft.com/en-us/training/support/mcp-developer-reference"},"s339":{"title":"xAI remote MCP tools","url":"https://docs.x.ai/developers/tools/remote-mcp"},"s340":{"title":"Streaming audio outputs","url":"https://developers.deepgram.com/docs/streaming-the-audio-output"},"s341":{"title":"Stream speech","url":"https://elevenlabs.io/docs/api-reference/text-to-speech/stream"},"s342":{"title":"Azure OpenAI text to speech quickstart","url":"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/text-to-speech-quickstart?view=foundry-classic"},"s344":{"title":"xAI video editing","url":"https://docs.x.ai/developers/model-capabilities/video/editing"},"s345":{"title":"xAI referencing files as Imagine input","url":"https://docs.x.ai/developers/model-capabilities/imagine/files/inputs"},"s346":{"title":"xAI video extension","url":"https://docs.x.ai/developers/model-capabilities/video/extension"},"s347":{"title":"xAI text to speech","url":"https://docs.x.ai/developers/model-capabilities/audio/text-to-speech"},"s351":{"title":"OpenAI compaction","url":"https://developers.openai.com/api/docs/guides/compaction"},"s352":{"title":"Get Image Generation","url":"https://elevenlabs.io/docs/api-reference/flows/image/get"},"s353":{"title":"Image & Video quickstart","url":"https://elevenlabs.io/docs/eleven-api/guides/cookbooks/image-and-video"},"s354":{"title":"References and assets","url":"https://elevenlabs.io/docs/eleven-api/guides/how-to/image-and-video/references"},"s355":{"title":"Get Video Generation","url":"https://elevenlabs.io/docs/api-reference/flows/video/get"},"s356":{"title":"Upload Asset","url":"https://elevenlabs.io/docs/api-reference/assets/create"},"s357":{"title":"Get Asset","url":"https://elevenlabs.io/docs/api-reference/assets/get"},"s358":{"title":"xAI streaming transcription schema","url":"https://docs.x.ai/stt-streaming.ws.json"},"s359":{"title":"Transcripts and commit strategies","url":"https://elevenlabs.io/docs/eleven-api/guides/how-to/speech-to-text/realtime/transcripts-and-commit-strategies"},"s360":{"title":"Mistral Libraries","url":"https://docs.mistral.ai/studio/search/libraries"},"s362":{"title":"Mistral human-in-the-loop confirmation","url":"https://docs.mistral.ai/studio/connectors/confirmation"},"s365":{"title":"Grounding with Vertex AI Search","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/grounding-with-vertex-ai-search"},"s366":{"title":"xAI speech to speech","url":"https://docs.x.ai/developers/model-capabilities/audio/speech-to-speech"},"s367":{"title":"xAI realtime WebSocket schema","url":"https://docs.x.ai/voice-realtime.ws.json"},"s368":{"title":"xAI WebSocket mode","url":"https://docs.x.ai/developers/advanced-api-usage/websocket-mode"},"s369":{"title":"Cohere create dataset","url":"https://docs.cohere.com/reference/create-dataset"},"s370":{"title":"Cohere get dataset","url":"https://docs.cohere.com/reference/get-dataset"},"s371":{"title":"Cohere datasets guide","url":"https://docs.cohere.com/docs/datasets"},"s372":{"title":"Build a Voice Agent","url":"https://developers.deepgram.com/reference/voice-agent/voice-agent"},"s374":{"title":"Agent WebSockets","url":"https://elevenlabs.io/docs/eleven-agents/api-reference/eleven-agents/websocket"},"s375":{"title":"Agent API custom functions","url":"https://docs.perplexity.ai/docs/agent-api/tools/custom-functions"},"s377":{"title":"Agent API sandbox","url":"https://docs.perplexity.ai/docs/agent-api/tools/sandbox"},"s378":{"title":"Agent API MCP","url":"https://docs.perplexity.ai/docs/agent-api/tools/mcp"},"s380":{"title":"Apply Patch server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/apply-patch"},"s381":{"title":"Upload a file to an active conversation","url":"https://elevenlabs.io/docs/api-reference/conversations/upload-file"},"s382":{"title":"Stored conversation usage, cost and transcript","url":"https://elevenlabs.io/docs/api-reference/conversations/get"},"s383":{"title":"Ollama commit removing image generation","url":"https://github.com/ollama/ollama/commit/4713800b08b2ddf5e14acf8398953cf7b12f169b"},"s385":{"title":"Source attribution defaults off","url":"https://elevenlabs.io/docs/api-reference/agents/create"},"s386":{"title":"Changelog, April 27, 2026","url":"https://elevenlabs.io/docs/changelog/2026/4/27"},"s396":{"title":"Cohere Azure workshops","url":"https://github.com/cohere-ai/cohere-azure-workshops"},"s397":{"title":"Cohere Embed v2 API","url":"https://docs.cohere.com/v2/reference/embed"},"s398":{"title":"Rank and rerank documents","url":"https://docs.cloud.google.com/generative-ai-app-builder/docs/ranking"},"s399":{"title":"rankingConfigs.rank reference","url":"https://docs.cloud.google.com/generative-ai-app-builder/docs/reference/rest/v1/projects.locations.rankingConfigs/rank"},"s400":{"title":"Luma Ray 2 parameters","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-luma.html"},"s401":{"title":"StartAsyncInvoke API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_StartAsyncInvoke.html"},"s402":{"title":"GetAsyncInvoke API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_GetAsyncInvoke.html"},"s404":{"title":"Batch inference models","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference-supported.html"},"s405":{"title":"Format batch inference data","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference-data.html"},"s406":{"title":"CreateModelInvocationJob API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_CreateModelInvocationJob.html"},"s410":{"title":"Gemini 3.5 Transcribe","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-5-transcribe"},"s412":{"title":"Orchestrator Chat Completions reference","url":"https://docs.perplexity.ai/api-reference/gateway-chat-completions-post"},"s413":{"title":"Orchestrator Messages reference","url":"https://docs.perplexity.ai/api-reference/gateway-messages-post"},"s414":{"title":"Batch embeddings","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/batch-prediction-genai-embeddings"},"s419":{"title":"ElevenLabs OpenAPI specification","url":"https://api.elevenlabs.io/openapi.json"},"s420":{"title":"eleven exa","url":"https://elevenlabs.io/docs/eleven-agents/customization/integrations/exa"},"s421":{"title":"eleven override","url":"https://elevenlabs.io/docs/eleven-agents/customization/personalization/overrides"},"s422":{"title":"RetrieveAndGenerate API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agent-runtime_RetrieveAndGenerate.html"},"s429":{"title":"ApplyGuardrail API reference","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ApplyGuardrail.html"},"s434":{"title":"Azure model availability by deployment type","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability?pivots=batch"},"s435":{"title":"Azure v1 Files REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/files"},"s439":{"title":"OpenAI Responses create reference","url":"https://developers.openai.com/api/reference/resources/responses/methods/create"},"s440":{"title":"OpenAI conversation state","url":"https://developers.openai.com/api/docs/guides/conversation-state"},"s441":{"title":"OpenAI streaming responses","url":"https://developers.openai.com/api/docs/guides/streaming-responses"},"s442":{"title":"OpenAI structured outputs","url":"https://developers.openai.com/api/docs/guides/structured-outputs"},"s443":{"title":"OpenAI images and vision","url":"https://developers.openai.com/api/docs/guides/images-vision"},"s444":{"title":"OpenAI file inputs","url":"https://developers.openai.com/api/docs/guides/file-inputs"},"s445":{"title":"OpenAI audio in Chat Completions","url":"https://developers.openai.com/api/docs/guides/audio-chat-completions"},"s446":{"title":"OpenAI deprecations","url":"https://developers.openai.com/api/docs/deprecations"},"s447":{"title":"OpenAI reasoning models","url":"https://developers.openai.com/api/docs/guides/reasoning"},"s448":{"title":"OpenAI function calling","url":"https://developers.openai.com/api/docs/guides/function-calling"},"s449":{"title":"OpenAI Code Interpreter","url":"https://developers.openai.com/api/docs/guides/tools-code-interpreter"},"s450":{"title":"OpenAI API reference index","url":"https://developers.openai.com/api/reference/llms.txt"},"s451":{"title":"OpenAI image generation","url":"https://developers.openai.com/api/docs/guides/image-generation"},"s452":{"title":"OpenAI Images reference","url":"https://developers.openai.com/api/reference/resources/images"},"s453":{"title":"OpenAI Videos reference","url":"https://developers.openai.com/api/reference/resources/videos"},"s454":{"title":"OpenAI Audio reference","url":"https://developers.openai.com/api/reference/resources/audio"},"s455":{"title":"OpenAI transcription","url":"https://developers.openai.com/api/docs/guides/transcription"},"s456":{"title":"OpenAI Realtime transcription","url":"https://developers.openai.com/api/docs/guides/realtime-transcription"},"s457":{"title":"OpenAI moderation","url":"https://developers.openai.com/api/docs/guides/moderation"},"s458":{"title":"OpenAI Moderations reference","url":"https://developers.openai.com/api/reference/resources/moderations"},"s459":{"title":"OpenAI API changelog","url":"https://developers.openai.com/api/docs/changelog"},"s460":{"title":"OpenAI vector embeddings","url":"https://developers.openai.com/api/docs/guides/embeddings"},"s461":{"title":"OpenAI Embeddings reference","url":"https://developers.openai.com/api/reference/resources/embeddings"},"s462":{"title":"OpenAI Files reference","url":"https://developers.openai.com/api/reference/resources/files"},"s463":{"title":"OpenAI Batches reference","url":"https://developers.openai.com/api/reference/resources/batches"},"s464":{"title":"OpenAI Skills reference","url":"https://developers.openai.com/api/reference/resources/skills"},"s465":{"title":"OpenAI all models","url":"https://developers.openai.com/api/docs/models/all"},"s466":{"title":"OpenAI Decisions API","url":"https://developers.openai.com/api/docs/guides/decisions"},"s467":{"title":"OpenAI Decisions reference","url":"https://developers.openai.com/api/reference/resources/decisions"},"s468":{"title":"OpenAI Ultrafast mode","url":"https://developers.openai.com/api/docs/guides/ultrafast-mode"},"s469":{"title":"OpenAI Vector Stores reference","url":"https://developers.openai.com/api/reference/resources/vector_stores"},"s470":{"title":"OpenAI Completions create reference","url":"https://developers.openai.com/api/reference/resources/completions/methods/create"},"s471":{"title":"OpenAI Completions API","url":"https://developers.openai.com/api/docs/guides/completions"},"s472":{"title":"OpenAI supervised fine-tuning","url":"https://developers.openai.com/api/docs/guides/supervised-fine-tuning"},"s473":{"title":"OpenAI Fine Tuning reference","url":"https://developers.openai.com/api/reference/resources/fine_tuning"},"s474":{"title":"OpenAI Conversations reference","url":"https://developers.openai.com/api/reference/resources/conversations"},"s475":{"title":"OpenAI Responses compact reference","url":"https://developers.openai.com/api/reference/resources/responses/methods/compact"},"s476":{"title":"OpenAI Programmatic Tool Calling","url":"https://developers.openai.com/api/docs/guides/tools-programmatic-tool-calling"},"s477":{"title":"OpenAI Realtime API","url":"https://developers.openai.com/api/docs/guides/realtime"},"s478":{"title":"OpenAI GPT-Live","url":"https://developers.openai.com/api/docs/guides/live"},"s479":{"title":"OpenAI migrate to Responses","url":"https://developers.openai.com/api/docs/guides/migrate-to-responses"},"s480":{"title":"OpenAI image generation tool","url":"https://developers.openai.com/api/docs/guides/tools-image-generation"},"s481":{"title":"OpenAI shell tool","url":"https://developers.openai.com/api/docs/guides/tools-shell"},"s482":{"title":"OpenAI custom voices","url":"https://developers.openai.com/api/docs/guides/custom-voices"},"s483":{"title":"OpenAI WebSocket mode","url":"https://developers.openai.com/api/docs/guides/websocket-mode"},"s484":{"title":"Create a Message","url":"https://platform.claude.com/docs/en/api/messages/create"},"s485":{"title":"Streaming messages","url":"https://platform.claude.com/docs/en/build-with-claude/streaming"},"s486":{"title":"Vision","url":"https://platform.claude.com/docs/en/build-with-claude/vision"},"s487":{"title":"PDF support","url":"https://platform.claude.com/docs/en/build-with-claude/pdf-support"},"s488":{"title":"Claude thinking","url":"https://platform.claude.com/docs/en/build-with-claude/thinking"},"s489":{"title":"Effort","url":"https://platform.claude.com/docs/en/build-with-claude/effort"},"s490":{"title":"Define tools","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools"},"s491":{"title":"Claude Platform release notes","url":"https://platform.claude.com/docs/en/release-notes/overview"},"s492":{"title":"Claude parallel tool use","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/parallel-tool-use"},"s493":{"title":"Claude web search tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool"},"s494":{"title":"Claude web fetch tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool"},"s495":{"title":"Claude code execution tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool"},"s496":{"title":"Claude threshold compaction","url":"https://platform.claude.com/docs/en/build-with-claude/compaction-threshold"},"s497":{"title":"Token counting","url":"https://platform.claude.com/docs/en/build-with-claude/token-counting"},"s498":{"title":"Count tokens in a Message","url":"https://platform.claude.com/docs/en/api/messages/count_tokens"},"s499":{"title":"Batch processing","url":"https://platform.claude.com/docs/en/build-with-claude/batch-processing"},"s500":{"title":"Claude advisor tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool"},"s501":{"title":"Skills in the API","url":"https://platform.claude.com/docs/en/build-with-claude/skills-guide"},"s502":{"title":"Claude browser use tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/browser-use-tool"},"s503":{"title":"Claude computer use tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool"},"s504":{"title":"Claude context editing","url":"https://platform.claude.com/docs/en/build-with-claude/context-editing"},"s505":{"title":"Claude refusals and fallback","url":"https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback"},"s506":{"title":"Message threads","url":"https://platform.claude.com/docs/en/build-with-claude/message-threads"},"s507":{"title":"Managed Agents sessions","url":"https://platform.claude.com/docs/en/managed-agents/sessions"},"s508":{"title":"Claude compaction on demand","url":"https://platform.claude.com/docs/en/build-with-claude/compaction-on-demand"},"s509":{"title":"Using the Messages API","url":"https://platform.claude.com/docs/en/build-with-claude/working-with-messages"},"s510":{"title":"Claude programmatic tool calling","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/programmatic-tool-calling"},"s511":{"title":"Claude bash tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/bash-tool"},"s512":{"title":"Claude text editor tool","url":"https://platform.claude.com/docs/en/agents-and-tools/tool-use/text-editor-tool"},"s513":{"title":"Gemini API overview","url":"https://aistudio.google.com/docs"},"s514":{"title":"Gemini text generation","url":"https://aistudio.google.com/docs/text-generation"},"s515":{"title":"GenerateContent API reference","url":"https://aistudio.google.com/api-reference/generate-content"},"s516":{"title":"Gemini Interactions API reference","url":"https://aistudio.google.com/api-reference/interactions-api"},"s517":{"title":"What's new in Gemini 3.6 Flash and 3.5 Flash-Lite","url":"https://aistudio.google.com/docs/whats-new-gemini-3.6"},"s518":{"title":"Gemini streaming","url":"https://aistudio.google.com/docs/streaming"},"s519":{"title":"Gemini structured outputs","url":"https://aistudio.google.com/docs/structured-output"},"s520":{"title":"Gemini image understanding","url":"https://aistudio.google.com/docs/image-understanding"},"s521":{"title":"Gemini document processing","url":"https://aistudio.google.com/docs/document-processing"},"s522":{"title":"Gemini audio understanding","url":"https://aistudio.google.com/docs/audio"},"s523":{"title":"Gemini video understanding","url":"https://aistudio.google.com/docs/video-understanding"},"s524":{"title":"Gemini thinking","url":"https://aistudio.google.com/docs/thinking"},"s525":{"title":"Gemini Google Search grounding","url":"https://aistudio.google.com/docs/google-search"},"s526":{"title":"Gemini function calling","url":"https://aistudio.google.com/docs/function-calling"},"s527":{"title":"Gemini tools overview","url":"https://aistudio.google.com/docs/tools"},"s528":{"title":"Gemini URL context","url":"https://aistudio.google.com/docs/url-context"},"s529":{"title":"Gemini code execution","url":"https://aistudio.google.com/docs/code-execution"},"s530":{"title":"Gemini File Search","url":"https://aistudio.google.com/docs/file-search"},"s531":{"title":"Gemini context caching","url":"https://aistudio.google.com/docs/caching"},"s532":{"title":"Gemini caching API reference","url":"https://aistudio.google.com/api-reference/caching"},"s533":{"title":"Gemini API all methods","url":"https://aistudio.google.com/api-reference/all-methods"},"s534":{"title":"Gemini Live session management","url":"https://aistudio.google.com/docs/live-api/session-management"},"s535":{"title":"Gemini Live API reference","url":"https://aistudio.google.com/api-reference/live"},"s536":{"title":"Gemini token counting","url":"https://aistudio.google.com/docs/tokens"},"s537":{"title":"Gemini counting tokens API reference","url":"https://aistudio.google.com/api-reference/tokens"},"s538":{"title":"Gemini Nano Banana image generation","url":"https://aistudio.google.com/docs/image-generation"},"s539":{"title":"Gemini Imagen","url":"https://aistudio.google.com/docs/imagen"},"s540":{"title":"Gemini Nano Banana 2.1 model","url":"https://aistudio.google.com/docs/models/gemini-nano-banana-2.1"},"s541":{"title":"Gemini deprecations","url":"https://aistudio.google.com/docs/deprecations"},"s542":{"title":"Gemini video generation","url":"https://aistudio.google.com/docs/video"},"s543":{"title":"Gemini Veo","url":"https://aistudio.google.com/docs/veo"},"s544":{"title":"Gemini Omni Flash","url":"https://aistudio.google.com/docs/omni"},"s545":{"title":"Gemini text-to-speech","url":"https://aistudio.google.com/docs/speech-generation"},"s546":{"title":"Gemini 3.8 Flash TTS model","url":"https://aistudio.google.com/docs/models/gemini-3.8-flash-tts"},"s547":{"title":"Gemini audio transcription","url":"https://aistudio.google.com/docs/transcribe"},"s548":{"title":"Gemini Live transcription","url":"https://aistudio.google.com/docs/live-api/live-transcribe"},"s549":{"title":"Gemini API reference","url":"https://aistudio.google.com/api-reference"},"s550":{"title":"Gemini embeddings","url":"https://aistudio.google.com/docs/embeddings"},"s551":{"title":"Gemini embeddings API reference","url":"https://aistudio.google.com/api-reference/embeddings"},"s552":{"title":"Gemini Files API","url":"https://aistudio.google.com/docs/files"},"s553":{"title":"Gemini files API reference","url":"https://aistudio.google.com/api-reference/files"},"s554":{"title":"Gemini Batch API","url":"https://aistudio.google.com/docs/batch-api"},"s555":{"title":"Gemini Batch API reference","url":"https://aistudio.google.com/api-reference/batch-api"},"s556":{"title":"Building Gemini managed agents","url":"https://aistudio.google.com/docs/custom-agents"},"s557":{"title":"Gemini agents overview","url":"https://aistudio.google.com/docs/agents"},"s558":{"title":"Gemini Deep Research agent","url":"https://aistudio.google.com/docs/deep-research"},"s559":{"title":"Gemini background execution","url":"https://aistudio.google.com/docs/background-execution"},"s560":{"title":"Gemini OpenAI compatibility","url":"https://aistudio.google.com/docs/openai"},"s561":{"title":"Gemini Computer Use","url":"https://aistudio.google.com/docs/computer-use"},"s562":{"title":"Gemini API release notes","url":"https://aistudio.google.com/docs/changelog"},"s563":{"title":"Gemini Live API","url":"https://aistudio.google.com/docs/live-api"},"s564":{"title":"Gemini Priority inference","url":"https://aistudio.google.com/docs/priority-inference"},"s565":{"title":"Gemini Flex inference","url":"https://aistudio.google.com/docs/flex-inference"},"s566":{"title":"Gemini fine-tuning","url":"https://aistudio.google.com/docs/model-tuning"},"s567":{"title":"Gemini Google Maps grounding","url":"https://aistudio.google.com/docs/maps-grounding"},"s568":{"title":"Gemini safety settings","url":"https://aistudio.google.com/docs/safety-settings"},"s569":{"title":"Gemini Interactions overview","url":"https://aistudio.google.com/docs/interactions-overview"},"s570":{"title":"Gemini Interactions OpenAPI","url":"https://aistudio.google.com/api-reference/interactions.openapi.json"},"s571":{"title":"Gemini Lyria music generation","url":"https://aistudio.google.com/docs/music-generation"},"s572":{"title":"Lyria 3.5 model","url":"https://aistudio.google.com/docs/models/lyria-3.5"},"s573":{"title":"Gemini Lyria RealTime","url":"https://aistudio.google.com/docs/realtime-music-generation"},"s574":{"title":"Gemini 2.5 Pro TTS model","url":"https://aistudio.google.com/docs/models/gemini-2.5-pro-preview-tts"},"s575":{"title":"Gemini Omni Flash model","url":"https://aistudio.google.com/docs/models/gemini-omni-flash"},"s576":{"title":"What's new in Gemini 3.8 Flash","url":"https://aistudio.google.com/docs/whats-new-gemini-3.8"},"s577":{"title":"Gemini tool combination","url":"https://aistudio.google.com/docs/tool-combination"},"s578":{"title":"Gemini Live capabilities","url":"https://aistudio.google.com/docs/live-api/capabilities"},"s579":{"title":"Gemini voice replication","url":"https://aistudio.google.com/docs/voice-replication"},"s580":{"title":"Gemini Voices API reference","url":"https://aistudio.google.com/api-reference/voices"},"s581":{"title":"Gemini voice remixing","url":"https://aistudio.google.com/docs/voice-remixing"},"s582":{"title":"Agent Platform beginners guide","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/beginners-guide"},"s583":{"title":"Send chat prompts to Gemini","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/send-chat-prompts-gemini"},"s584":{"title":"Use Claude models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/use-claude"},"s585":{"title":"Use partner models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/use-partner-models"},"s586":{"title":"Call open model APIs","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/call-open-model-apis"},"s587":{"title":"Control generated output","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/control-generated-output"},"s588":{"title":"Claude structured outputs","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/structured-outputs"},"s589":{"title":"Open model structured output","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/capabilities/structured-output"},"s590":{"title":"Image understanding","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/image-understanding"},"s591":{"title":"Claude Models on Google Cloud (Google managed)","url":"https://platform.claude.com/docs/en/build-with-claude/claude-models-on-google-cloud/google-managed"},"s592":{"title":"Document understanding","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/document-understanding"},"s593":{"title":"Audio understanding","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/audio-understanding"},"s594":{"title":"Video understanding","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/video-understanding"},"s595":{"title":"Gemini thinking","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/thinking"},"s596":{"title":"Open model thinking","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/capabilities/thinking"},"s597":{"title":"Grounding with Google Search","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/grounding-with-google-search"},"s598":{"title":"Gemini function calling","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tools/function-calling"},"s599":{"title":"Open model function calling","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/capabilities/function-calling"},"s600":{"title":"Grok function calling","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/grok/capabilities/function-calling"},"s601":{"title":"Claude web search","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/web-search"},"s602":{"title":"URL context","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/url-context"},"s603":{"title":"Claude Google managed and Anthropic managed comparison","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/anthropic-managed"},"s604":{"title":"Code execution API","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/code-execution-api"},"s605":{"title":"Claude prompt caching","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/prompt-caching"},"s606":{"title":"Count tokens for Claude models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/count-tokens"},"s607":{"title":"Gemini 3.8 Flash TTS","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash-tts"},"s608":{"title":"Gemini 3.8 Flash","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash"},"s609":{"title":"Generate images with Gemini","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/image-generation"},"s610":{"title":"Google models on Agent Platform","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models"},"s611":{"title":"Gemini 3 Pro Image","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-pro-image"},"s612":{"title":"Imagen subject customization (discontinued endpoints)","url":"https://docs.cloud.google.com/vertex-ai/generative-ai/docs/image/subject-customization"},"s613":{"title":"Edit images with Gemini","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/gemini-edit-images"},"s614":{"title":"Veo 3.1 model reference","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate"},"s615":{"title":"Generate speech with Gemini TTS","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/text-to-speech/overview"},"s616":{"title":"Cloud Text-to-Speech REST reference","url":"https://docs.cloud.google.com/text-to-speech/docs/reference/rest"},"s617":{"title":"Gemini for safety filtering and content moderation","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/gemini-for-filtering-and-moderation"},"s618":{"title":"Safety overview","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/safety-overview"},"s619":{"title":"Agent Platform APIs","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/apis"},"s620":{"title":"Mistral OCR (25.05)","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/mistral/mistral-ocr"},"s621":{"title":"DeepSeek-OCR","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/deepseek/deepseek-ocr"},"s622":{"title":"Get text embeddings","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/embeddings/get-text-embeddings"},"s623":{"title":"Cloud Storage objects API","url":"https://docs.cloud.google.com/storage/docs/json_api/v1/objects"},"s624":{"title":"Batch inference with Gemini","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/batch-inference"},"s625":{"title":"Batch predictions for Claude","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/batch"},"s626":{"title":"Batch predictions for open models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/maas/capabilities/batch-prediction"},"s627":{"title":"Agent Platform REST API reference","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/rest"},"s628":{"title":"Interactions API overview","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/interactions"},"s629":{"title":"Grok 4.7","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/grok/grok-4-7"},"s630":{"title":"Gemini computer use","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/computer-use"},"s631":{"title":"Agent Platform release notes","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes"},"s632":{"title":"Flex PayGo","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/flex-paygo"},"s633":{"title":"Codestral 2","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/mistral/codestral-2"},"s634":{"title":"Mistral models on Agent Platform","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/mistral"},"s635":{"title":"Tuning API","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/reference/models/tuning"},"s636":{"title":"Tune Gemini models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning"},"s637":{"title":"Grounding with Google Maps","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/grounding-with-google-maps"},"s638":{"title":"Configure safety filters","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/configure-safety-filters"},"s639":{"title":"Use Llama models","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/llama/use-llama"},"s640":{"title":"Interactions API developer guide","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/interactions/developer-guide"},"s641":{"title":"Generate music with Lyria","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/music/generate-music"},"s642":{"title":"Lyria overview","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/music/overview"},"s643":{"title":"Gemini Omni 1.2 Flash","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/omni-1-2-flash"},"s644":{"title":"Claude on Agent Platform","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude"},"s645":{"title":"Voice replication","url":"https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/text-to-speech/voice-replication"},"s646":{"title":"APIs supported by Amazon Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/apis.html"},"s647":{"title":"Bedrock Converse API guide","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html"},"s648":{"title":"Bedrock Messages API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-messages-api.html"},"s649":{"title":"Claude in Amazon Bedrock","url":"https://platform.claude.com/docs/en/build-with-claude/claude-in-amazon-bedrock"},"s650":{"title":"Bedrock Invoke API guide","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-api.html"},"s651":{"title":"Bedrock Responses API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-responses-api.html"},"s652":{"title":"Bedrock structured outputs","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/structured-output.html"},"s653":{"title":"Structured outputs for Claude on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-structured-outputs.html"},"s654":{"title":"GPT-5.6 Terra model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-terra.html"},"s655":{"title":"Converse ContentBlock","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ContentBlock.html"},"s656":{"title":"Claude on Amazon Bedrock (Opus 4.6 and earlier)","url":"https://platform.claude.com/docs/en/build-with-claude/claude-on-amazon-bedrock-legacy"},"s657":{"title":"Bedrock Chat Completions API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-chat-completions.html"},"s658":{"title":"Voxtral Mini 3B model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-mistral-ai-voxtral-mini-3b-2507.html"},"s659":{"title":"Claude adaptive thinking on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-adaptive-thinking.html"},"s660":{"title":"Claude extended thinking on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html"},"s661":{"title":"Bedrock inference reasoning","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-reasoning.html"},"s662":{"title":"Bedrock client-side tool use","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/tool-use-client-side.html"},"s663":{"title":"Claude tool use on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages-tool-use.html"},"s664":{"title":"Converse ToolChoice","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ToolChoice.html"},"s665":{"title":"Nova 2 Web Grounding","url":"https://docs.aws.amazon.com/nova/latest/nova2-userguide/web-grounding.html"},"s666":{"title":"Nova 2 tool use and built-in system tools","url":"https://docs.aws.amazon.com/nova/latest/nova2-userguide/using-tools.html"},"s667":{"title":"GPT-5.4 model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-54.html"},"s668":{"title":"Claude compaction on Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-compaction.html"},"s669":{"title":"Bedrock CountTokens guide","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/count-tokens.html"},"s670":{"title":"Nova Canvas model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-canvas.html"},"s671":{"title":"Luma Ray 2 model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-luma-ai-ray2.html"},"s672":{"title":"Nova Reel model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-reel.html"},"s673":{"title":"Nova 2.5 Sonic model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-2-5-sonic.html"},"s674":{"title":"InvokeModelWithBidirectionalStream","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModelWithBidirectionalStream.html"},"s675":{"title":"Use the InvokeGuardrailChecks API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-use-invoke-guardrail-checks.html"},"s676":{"title":"Bedrock models at a glance","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html"},"s677":{"title":"Cohere Embed v4 model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-cohere-embed-v4.html"},"s678":{"title":"Nova Multimodal Embeddings model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-amazon-nova-multimodal-embeddings.html"},"s679":{"title":"Titan Multimodal Embeddings G1","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/titan-multiemb-models.html"},"s680":{"title":"Use a reranker model","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/rerank-use.html"},"s681":{"title":"Bedrock OpenAI Batch API","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/inference-openai-batch.html"},"s682":{"title":"Claude Agent Skills","url":"https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview"},"s683":{"title":"Inference endpoints supported by Amazon Bedrock","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/endpoints.html"},"s684":{"title":"Bedrock computer use","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/computer-use.html"},"s685":{"title":"Converse PerformanceConfiguration","url":"https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_PerformanceConfiguration.html"},"s686":{"title":"Bedrock service tiers","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html"},"s687":{"title":"Bedrock OpenAI-compatible fine-tuning","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/fine-tuning-openai-apis.html"},"s688":{"title":"Guardrails with inference operations","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-input-tagging-base-inference.html"},"s689":{"title":"GPT-6 Sol model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-sol.html"},"s690":{"title":"GPT OSS 20B model card","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-oss-20b.html"},"s691":{"title":"Azure OpenAI chat completions","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/chatgpt"},"s692":{"title":"Azure OpenAI Chat REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/chat"},"s693":{"title":"Azure OpenAI structured outputs","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/structured-outputs"},"s694":{"title":"Azure OpenAI vision-enabled chats","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/gpt-with-vision"},"s695":{"title":"Azure OpenAI audio generation","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/audio-completions-quickstart"},"s696":{"title":"Azure OpenAI Responses REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/responses"},"s697":{"title":"Azure OpenAI reasoning models","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning"},"s698":{"title":"Azure OpenAI function calling","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/function-calling"},"s699":{"title":"Azure vector stores REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/vector_stores"},"s700":{"title":"Microsoft Foundry REST API reference","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/"},"s701":{"title":"Claude models in Microsoft Foundry","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/claude-models"},"s702":{"title":"Azure OpenAI model retirement schedule","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule"},"s703":{"title":"Speech to text with transcription models","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/whisper-quickstart"},"s704":{"title":"Foundry Models sold by Azure","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure"},"s705":{"title":"Foundry REST: create a moderation","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/moderations-create-moderation"},"s706":{"title":"Azure OpenAI embeddings REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/embeddings"},"s707":{"title":"Foundry REST: create a batch","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/batches-create-batch"},"s708":{"title":"Azure Responses skills","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/skills"},"s709":{"title":"Azure Responses shell tool","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/shells"},"s710":{"title":"Azure deep research","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/deep-research"},"s711":{"title":"Azure OpenAI computer use","url":"https://learn.microsoft.com/en-us/azure/foundry-classic/openai/how-to/computer-use"},"s712":{"title":"Azure OpenAI Decisions API","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/decisions"},"s713":{"title":"Azure priority processing","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/priority-processing"},"s714":{"title":"Azure Flex processing","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/flex-processing"},"s715":{"title":"Foundry REST: create a completion","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/completions-create-completion"},"s716":{"title":"Foundry Models from partners and community","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-from-partners"},"s717":{"title":"Azure fine-tuning REST reference","url":"https://learn.microsoft.com/en-us/rest/api/microsoft-foundry/azureopenai/fine-tuning"},"s718":{"title":"Configure Foundry guardrails","url":"https://learn.microsoft.com/en-us/azure/foundry/guardrails/how-to-create-guardrails"},"s719":{"title":"Foundry REST: list conversations","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/list-conversations"},"s720":{"title":"Azure OpenAI JSON mode","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/json-mode"},"s721":{"title":"Deploy and use Claude models in Foundry","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-claude"},"s722":{"title":"Grok models in Foundry","url":"https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-grok"},"s723":{"title":"Azure programmatic tool calling","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/programmatic-tool-calling"},"s724":{"title":"Azure model router","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-router"},"s725":{"title":"Azure provisioned spillover","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/spillover-traffic-management"},"s726":{"title":"Azure GPT Realtime API","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/realtime-audio"},"s727":{"title":"Azure Responses tool search","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/tool-search"},"s728":{"title":"Foundry REST: create a video remix","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/video-generation-create-video-remix"},"s729":{"title":"Foundry REST: create a video edit","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/video-generation-create-video-edit-multipart-video-generation-create-video-edit-json"},"s730":{"title":"Foundry REST: extend a video","url":"https://learn.microsoft.com/en-us/azure/foundry/reference/rest/video-generation-create-video-extend-multipart-video-generation-create-video-extend-json"},"s731":{"title":"xAI text generation","url":"https://docs.x.ai/developers/model-capabilities/text/generate-text"},"s732":{"title":"xAI Chat Completions reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/chat-completions"},"s733":{"title":"xAI multi-agent","url":"https://docs.x.ai/developers/model-capabilities/text/multi-agent"},"s734":{"title":"xAI streaming","url":"https://docs.x.ai/developers/model-capabilities/text/streaming"},"s735":{"title":"xAI Responses API reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/responses"},"s736":{"title":"xAI structured outputs","url":"https://docs.x.ai/developers/model-capabilities/text/structured-outputs"},"s737":{"title":"xAI image understanding","url":"https://docs.x.ai/developers/model-capabilities/images/understanding"},"s738":{"title":"xAI chat with files","url":"https://docs.x.ai/developers/model-capabilities/files/chat-with-files"},"s739":{"title":"xAI OpenAPI specification","url":"https://docs.x.ai/openapi.json"},"s740":{"title":"xAI X search tool","url":"https://docs.x.ai/developers/tools/x-search"},"s741":{"title":"xAI reasoning","url":"https://docs.x.ai/developers/model-capabilities/text/reasoning"},"s742":{"title":"xAI citations","url":"https://docs.x.ai/developers/tools/citations"},"s743":{"title":"xAI collections search tool","url":"https://docs.x.ai/developers/tools/collections-search"},"s744":{"title":"xAI function calling","url":"https://docs.x.ai/developers/tools/function-calling"},"s745":{"title":"xAI web search tool","url":"https://docs.x.ai/developers/tools/web-search"},"s746":{"title":"xAI code execution tool","url":"https://docs.x.ai/developers/tools/code-execution"},"s747":{"title":"xAI Collections","url":"https://docs.x.ai/developers/files/collections"},"s748":{"title":"xAI maximizing cache hits","url":"https://docs.x.ai/developers/advanced-api-usage/prompt-caching/maximizing-cache-hits"},"s749":{"title":"xAI legacy and deprecated endpoints","url":"https://docs.x.ai/developers/rest-api-reference/inference/legacy"},"s750":{"title":"xAI image generation","url":"https://docs.x.ai/developers/model-capabilities/images/generation"},"s751":{"title":"xAI Images reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/images"},"s752":{"title":"xAI image editing","url":"https://docs.x.ai/developers/model-capabilities/images/editing"},"s753":{"title":"xAI multi-image editing","url":"https://docs.x.ai/developers/model-capabilities/images/multi-image-editing"},"s754":{"title":"xAI video generation","url":"https://docs.x.ai/developers/model-capabilities/video/generation"},"s755":{"title":"xAI reference-to-video","url":"https://docs.x.ai/developers/model-capabilities/video/reference-to-video"},"s756":{"title":"xAI REST API reference overview","url":"https://docs.x.ai/developers/rest-api-reference/inference"},"s757":{"title":"xAI custom moderation controls","url":"https://docs.x.ai/developers/advanced-api-usage/custom-moderation-controls"},"s758":{"title":"xAI Collections reference","url":"https://docs.x.ai/developers/rest-api-reference/collections/collection"},"s759":{"title":"xAI managing files","url":"https://docs.x.ai/developers/files/managing-files"},"s760":{"title":"xAI Files reference","url":"https://docs.x.ai/developers/rest-api-reference/files/manage"},"s761":{"title":"xAI Batches reference","url":"https://docs.x.ai/developers/rest-api-reference/inference/batches"},"s762":{"title":"xAI collection metadata","url":"https://docs.x.ai/developers/files/collections/metadata"},"s763":{"title":"xAI collections search reference","url":"https://docs.x.ai/developers/rest-api-reference/collections/search"},"s764":{"title":"xAI image generation tool","url":"https://docs.x.ai/developers/tools/image-generation"},"s765":{"title":"xAI tool usage details","url":"https://docs.x.ai/developers/tools/tool-usage-details"},"s766":{"title":"DeepSeek Chat Completions reference","url":"https://api-docs.deepseek.com/api/create-chat-completion"},"s767":{"title":"DeepSeek Models and Pricing","url":"https://api-docs.deepseek.com/quick_start/pricing"},"s768":{"title":"DeepSeek Vision guide","url":"https://api-docs.deepseek.com/guides/vision"},"s769":{"title":"DeepSeek Anthropic API guide","url":"https://api-docs.deepseek.com/guides/anthropic_api"},"s770":{"title":"DeepSeek Lists Models reference","url":"https://api-docs.deepseek.com/api/list-models"},"s771":{"title":"DeepSeek Thinking Mode guide","url":"https://api-docs.deepseek.com/guides/thinking_mode"},"s772":{"title":"DeepSeek Tool Calls guide","url":"https://api-docs.deepseek.com/guides/tool_calls"},"s773":{"title":"DeepSeek Context Caching guide","url":"https://api-docs.deepseek.com/guides/kv_cache"},"s774":{"title":"DeepSeek Change Log","url":"https://api-docs.deepseek.com/updates"},"s775":{"title":"DeepSeek Rate Limit and Isolation","url":"https://api-docs.deepseek.com/quick_start/rate_limit"},"s776":{"title":"DeepSeek Search and Fetch guide","url":"https://api-docs.deepseek.com/guides/search"},"s777":{"title":"DeepSeek Upload File reference","url":"https://api-docs.deepseek.com/api/create-file"},"s778":{"title":"DeepSeek Retrieve File reference","url":"https://api-docs.deepseek.com/api/retrieve-file"},"s779":{"title":"DeepSeek List Files reference","url":"https://api-docs.deepseek.com/api/list-files"},"s780":{"title":"DeepSeek Delete File reference","url":"https://api-docs.deepseek.com/api/delete-file"},"s781":{"title":"DeepSeek FIM Completion guide","url":"https://api-docs.deepseek.com/guides/fim_completion"},"s782":{"title":"DeepSeek FIM Completion reference","url":"https://api-docs.deepseek.com/api/create-completion"},"s783":{"title":"DeepSeek JSON Output guide","url":"https://api-docs.deepseek.com/guides/json_mode"},"s784":{"title":"DeepSeek Chat Prefix Completion guide","url":"https://api-docs.deepseek.com/guides/chat_prefix_completion"},"s785":{"title":"DeepSeek Search reference","url":"https://api-docs.deepseek.com/api/search"},"s786":{"title":"DeepSeek Fetch reference","url":"https://api-docs.deepseek.com/api/fetch"},"s787":{"title":"Cohere Chat guide","url":"https://docs.cohere.com/docs/chat-api"},"s788":{"title":"Cohere streaming guide","url":"https://docs.cohere.com/docs/streaming"},"s789":{"title":"Cohere structured outputs","url":"https://docs.cohere.com/docs/structured-outputs"},"s790":{"title":"Cohere image inputs","url":"https://docs.cohere.com/docs/image-inputs"},"s791":{"title":"Cohere RAG citations","url":"https://docs.cohere.com/docs/rag-citations"},"s792":{"title":"Cohere tool use overview","url":"https://docs.cohere.com/docs/tools"},"s793":{"title":"Cohere tool use patterns","url":"https://docs.cohere.com/docs/tool-use-usage-patterns"},"s794":{"title":"Cohere Compatibility API","url":"https://docs.cohere.com/docs/compatibility-api"},"s795":{"title":"Cohere v1 to v2 migration","url":"https://docs.cohere.com/docs/migrating-v1-to-v2"},"s796":{"title":"Cohere Chat API (v1)","url":"https://docs.cohere.com/reference/chat-v1"},"s797":{"title":"Cohere v2 API release notes","url":"https://docs.cohere.com/changelog/v2-api-release"},"s798":{"title":"Cohere tokens and tokenizers","url":"https://docs.cohere.com/docs/tokens-and-tokenizers"},"s799":{"title":"Cohere API reference overview","url":"https://docs.cohere.com/reference/about"},"s800":{"title":"Cohere release notes","url":"https://docs.cohere.com/changelog"},"s801":{"title":"Cohere Transcribe","url":"https://docs.cohere.com/docs/transcribe"},"s802":{"title":"Cohere Moderate beta notice","url":"https://docs.cohere.com/changelog/introducing-moderate-beta"},"s803":{"title":"Cohere Parse model","url":"https://docs.cohere.com/docs/parse"},"s804":{"title":"Cohere Embed API","url":"https://docs.cohere.com/reference/embed"},"s805":{"title":"Cohere multimodal embeddings","url":"https://docs.cohere.com/docs/multimodal-embeddings"},"s806":{"title":"Cohere Rerank models","url":"https://docs.cohere.com/docs/rerank"},"s807":{"title":"Cohere cancel batch","url":"https://docs.cohere.com/reference/cancel-batch"},"s808":{"title":"Cohere 2025-09-15 deprecation notice","url":"https://docs.cohere.com/changelog/2025-09-15-major-command-deprecations"},"s809":{"title":"Cohere safety modes","url":"https://docs.cohere.com/docs/safety-modes"},"s810":{"title":"Mistral chat completions","url":"https://docs.mistral.ai/inference/conversations/chat-completion"},"s811":{"title":"Mistral custom structured outputs","url":"https://docs.mistral.ai/inference/conversations/structured-output/custom"},"s812":{"title":"Mistral vision","url":"https://docs.mistral.ai/inference/conversations/vision"},"s813":{"title":"Mistral Document QnA","url":"https://docs.mistral.ai/studio/document-processing/document_qna"},"s814":{"title":"Mistral audio in chat completions","url":"https://docs.mistral.ai/studio/audio/audio-in-chat"},"s815":{"title":"Mistral OpenAPI specification","url":"https://docs.mistral.ai/openapi.yaml"},"s816":{"title":"Mistral reasoning","url":"https://docs.mistral.ai/inference/conversations/reasoning"},"s817":{"title":"Mistral citations and references","url":"https://docs.mistral.ai/inference/conversations/citations"},"s818":{"title":"Mistral function calling","url":"https://docs.mistral.ai/inference/conversations/function-calling"},"s819":{"title":"Mistral move from the Agents API","url":"https://docs.mistral.ai/studio/agents/getting-started/from-agents-api"},"s820":{"title":"Mistral code interpreter tool","url":"https://docs.mistral.ai/studio/agents/agent-tools/code_interpreter"},"s821":{"title":"Mistral connectors","url":"https://docs.mistral.ai/studio/connectors/overview"},"s822":{"title":"Mistral connectors in conversations","url":"https://docs.mistral.ai/studio/connectors/conversations"},"s823":{"title":"Mistral prompt caching","url":"https://docs.mistral.ai/inference/conversations/advanced/prompt-caching"},"s824":{"title":"Mistral Responses API guide","url":"https://docs.mistral.ai/inference/conversations/responses"},"s825":{"title":"Mistral documentation index","url":"https://docs.mistral.ai/llms.txt"},"s826":{"title":"Mistral Managed Agents","url":"https://docs.mistral.ai/studio/agents/overview"},"s827":{"title":"Mistral Responses API reference","url":"https://docs.mistral.ai/api/endpoint/responses"},"s828":{"title":"Mistral speech generation","url":"https://docs.mistral.ai/studio/audio/text-to-speech/generate-speech"},"s829":{"title":"Mistral file transcription","url":"https://docs.mistral.ai/studio/audio/speech-to-text/transcribe-files"},"s830":{"title":"Mistral realtime transcription","url":"https://docs.mistral.ai/studio/audio/speech-to-text/transcribe-in-realtime"},"s831":{"title":"Mistral speaker diarization","url":"https://docs.mistral.ai/studio/audio/speech-to-text/diarize-speakers"},"s832":{"title":"Mistral moderation and guardrails","url":"https://docs.mistral.ai/studio/safety-moderation"},"s833":{"title":"Mistral OCR processor","url":"https://docs.mistral.ai/studio/document-processing/basic_ocr"},"s834":{"title":"Mistral text embeddings","url":"https://docs.mistral.ai/inference/embeddings/text_embeddings"},"s835":{"title":"Mistral batch processing","url":"https://docs.mistral.ai/inference/batch-processing"},"s836":{"title":"Mistral Batch API","url":"https://docs.mistral.ai/api/endpoint/batch"},"s837":{"title":"Mistral Prompts and Skills","url":"https://docs.mistral.ai/studio/prompts-and-skills/overview"},"s838":{"title":"Mistral Classifiers API","url":"https://docs.mistral.ai/api/endpoint/classifiers"},"s839":{"title":"Mistral Priority Tier","url":"https://docs.mistral.ai/inference/priority-tier"},"s840":{"title":"Mistral FIM API","url":"https://docs.mistral.ai/api/endpoint/fim"},"s841":{"title":"Mistral changelog","url":"https://docs.mistral.ai/resources/changelogs"},"s842":{"title":"Mistral JSON mode","url":"https://docs.mistral.ai/inference/conversations/structured-output/json_mode"},"s843":{"title":"Mistral voice agents","url":"https://docs.mistral.ai/studio/audio/voice-agents/overview"},"s844":{"title":"Mistral realtime speech","url":"https://docs.mistral.ai/studio/audio/text-to-speech/speak-in-realtime"},"s845":{"title":"Mistral Realtime Sessions API","url":"https://docs.mistral.ai/api/endpoint/realtime/sessions"},"s846":{"title":"Mistral Search Index","url":"https://docs.mistral.ai/studio/search/search-index/overview"},"s847":{"title":"Mistral stream speech","url":"https://docs.mistral.ai/studio/audio/text-to-speech/stream-speech"},"s848":{"title":"Mistral Search Index API","url":"https://docs.mistral.ai/api/endpoint/beta/context/search_indexes"},"s849":{"title":"Mistral voice cloning","url":"https://docs.mistral.ai/studio/audio/text-to-speech/voice-cloning"},"s850":{"title":"Mistral Agents and Conversations","url":"https://docs.mistral.ai/studio/agents/agents-api"},"s851":{"title":"API reference overview","url":"https://openrouter.ai/docs/api_reference/overview"},"s852":{"title":"Streaming","url":"https://openrouter.ai/docs/api_reference/streaming"},"s853":{"title":"Structured Outputs","url":"https://openrouter.ai/docs/guides/features/structured-outputs"},"s854":{"title":"API parameters","url":"https://openrouter.ai/docs/api_reference/parameters"},"s855":{"title":"Image Inputs","url":"https://openrouter.ai/docs/guides/overview/multimodal/image-understanding"},"s856":{"title":"OpenRouter OpenAPI specification","url":"https://openrouter.ai/openapi.json"},"s857":{"title":"Video Inputs","url":"https://openrouter.ai/docs/guides/overview/multimodal/videos"},"s858":{"title":"Reasoning Tokens","url":"https://openrouter.ai/docs/guides/best-practices/reasoning-tokens"},"s859":{"title":"Web Search server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/web-search"},"s860":{"title":"Client Tools","url":"https://openrouter.ai/docs/guides/features/tool-calling"},"s861":{"title":"Web Fetch server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/web-fetch"},"s862":{"title":"Create transcription","url":"https://openrouter.ai/docs/api/api-reference/stt/create-transcription"},"s863":{"title":"Guardrail Policies","url":"https://openrouter.ai/docs/guides/features/guardrails/policies"},"s864":{"title":"Extractions API","url":"https://openrouter.ai/docs/guides/community/extractions"},"s865":{"title":"Submit an embedding request","url":"https://openrouter.ai/docs/api/api-reference/embeddings/submit-an-embedding-request"},"s866":{"title":"Submit a rerank request","url":"https://openrouter.ai/docs/api/api-reference/rerank/submit-a-rerank-request"},"s867":{"title":"Batch API","url":"https://openrouter.ai/docs/guides/features/batch-api"},"s868":{"title":"Advisor server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/advisor"},"s869":{"title":"Create a message","url":"https://openrouter.ai/docs/api/api-reference/anthropic-messages/create-a-message"},"s870":{"title":"Submit a Decisions request","url":"https://openrouter.ai/docs/api/api-reference/alphadecisions/submit-a-decisions-request"},"s871":{"title":"Classify and Tag Text at Scale with Jev","url":"https://openrouter.ai/docs/cookbook/evaluate-and-optimize/jev-classification"},"s872":{"title":"List files","url":"https://openrouter.ai/docs/api/api-reference/files/list-files"},"s873":{"title":"List guardrails","url":"https://openrouter.ai/docs/api/api-reference/guardrails/list-guardrails"},"s874":{"title":"Claude 5.5 Migration Guide","url":"https://openrouter.ai/docs/cookbook/evaluate-and-optimize/model-migrations/claude-5-5"},"s875":{"title":"Image Generation server tool","url":"https://openrouter.ai/docs/guides/features/server-tools/image-generation"},"s876":{"title":"Codex Desktop App","url":"https://openrouter.ai/docs/cookbook/coding-agents/codex-desktop"},"s877":{"title":"Migrate from Sonar to the Agent API","url":"https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview"},"s878":{"title":"Agent API output control","url":"https://docs.perplexity.ai/docs/agent-api/output-control"},"s879":{"title":"Agent API image attachments","url":"https://docs.perplexity.ai/docs/agent-api/image-attachments"},"s880":{"title":"Sonar media and attachments (legacy)","url":"https://docs.perplexity.ai/docs/legacy/sonar/media"},"s881":{"title":"Agent API working with files","url":"https://docs.perplexity.ai/docs/agent-api/working-with-files"},"s882":{"title":"Perplexity API changelog","url":"https://docs.perplexity.ai/docs/resources/changelog"},"s883":{"title":"How to migrate from Sonar","url":"https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/how-to"},"s884":{"title":"Sonar Deep Research (legacy)","url":"https://docs.perplexity.ai/docs/legacy/sonar/models/sonar-deep-research"},"s885":{"title":"Agent API effort levels","url":"https://docs.perplexity.ai/docs/agent-api/effort"},"s886":{"title":"Agent API connectors","url":"https://docs.perplexity.ai/docs/agent-api/tools/connectors"},"s887":{"title":"Agent API Batch API","url":"https://docs.perplexity.ai/docs/agent-api/batch"},"s888":{"title":"Agent API background mode","url":"https://docs.perplexity.ai/docs/agent-api/background-mode"},"s889":{"title":"Perplexity Decisions API","url":"https://docs.perplexity.ai/docs/decisions/quickstart"},"s890":{"title":"Perplexity contextualized embeddings","url":"https://docs.perplexity.ai/docs/embeddings/contextualized-embeddings"},"s891":{"title":"Agent API models and service tiers","url":"https://docs.perplexity.ai/docs/agent-api/models"},"s892":{"title":"Agent API model fallback","url":"https://docs.perplexity.ai/docs/agent-api/model-fallback"},"s893":{"title":"Perplexity Search API overview","url":"https://docs.perplexity.ai/docs/search/overview"},"s894":{"title":"Agent API define the run","url":"https://docs.perplexity.ai/docs/agent-api/building-agents/define-the-run"},"s895":{"title":"Deepgram LLMs overview","url":"https://developers.deepgram.com/docs/llm"},"s896":{"title":"Deepgram LLMs request and response format","url":"https://developers.deepgram.com/docs/llm/request-response-format"},"s897":{"title":"Voice Agent Agent Thinking message","url":"https://developers.deepgram.com/docs/voice-agent-agent-thinking"},"s898":{"title":"Deepgram LLMs models overview","url":"https://developers.deepgram.com/docs/llm/models"},"s899":{"title":"Use Deepgram LLMs in the Voice Agent","url":"https://developers.deepgram.com/docs/llm/voice-agent"},"s900":{"title":"Voice Agent Function Call Request","url":"https://developers.deepgram.com/docs/voice-agent-function-call-request"},"s901":{"title":"Flux TTS batch (REST)","url":"https://developers.deepgram.com/docs/flux-tts/batch"},"s902":{"title":"Migrating from /v1/speak to Flux TTS","url":"https://developers.deepgram.com/docs/flux-tts/migrating"},"s903":{"title":"Flux Text to Speech (batch) reference","url":"https://developers.deepgram.com/reference/speak/v-2/audio/generate"},"s904":{"title":"Turn-based audio (Flux) reference","url":"https://developers.deepgram.com/reference/speech-to-text/listen-flux"},"s905":{"title":"Analyze Text reference","url":"https://developers.deepgram.com/reference/text-intelligence/analyze-text"},"s906":{"title":"Voice Agent LLM models","url":"https://developers.deepgram.com/docs/voice-agent-llm-models"},"s907":{"title":"Turn-based Speech (Flux) reference","url":"https://developers.deepgram.com/reference/text-to-speech/speak-flux"},"s908":{"title":"ElevenLabs changelog","url":"https://elevenlabs.io/docs/changelog"},"s909":{"title":"Create realtime client secret","url":"https://elevenlabs.io/docs/eleven-agents/api-reference/conversations/create-client-secret"},"s910":{"title":"OpenAI Realtime API compatibility","url":"https://elevenlabs.io/docs/eleven-agents/guides/openai-realtime-compatibility"},"s911":{"title":"Realtime models","url":"https://elevenlabs.io/docs/eleven-agents/customization/realtime-models"},"s912":{"title":"Create speech","url":"https://elevenlabs.io/docs/api-reference/text-to-speech/convert"},"s913":{"title":"Asynchronous Speech to Text","url":"https://elevenlabs.io/docs/eleven-api/guides/how-to/speech-to-text/batch/webhooks"},"s914":{"title":"Dub a video or an audio file","url":"https://elevenlabs.io/docs/api-reference/dubbing/create"},"s915":{"title":"Dubbing","url":"https://elevenlabs.io/docs/overview/capabilities/dubbing"},"s916":{"title":"List Assets","url":"https://elevenlabs.io/docs/api-reference/assets/list"},"s917":{"title":"Delete Asset","url":"https://elevenlabs.io/docs/api-reference/assets/delete"},"s918":{"title":"Eleven Music","url":"https://elevenlabs.io/docs/overview/capabilities/music"},"s919":{"title":"LLM Cascading","url":"https://elevenlabs.io/docs/eleven-agents/customization/llm/llm-cascading"},"s920":{"title":"Create sound effect","url":"https://elevenlabs.io/docs/api-reference/text-to-sound-effects/convert"},"s921":{"title":"Sound effects","url":"https://elevenlabs.io/docs/overview/capabilities/sound-effects"},"s922":{"title":"WebSocket text to speech","url":"https://elevenlabs.io/docs/api-reference/text-to-speech/v-1-text-to-speech-voice-id-stream-input"},"s923":{"title":"Create IVC voice","url":"https://elevenlabs.io/docs/api-reference/voices/ivc/create"},"s924":{"title":"Create PVC voice","url":"https://elevenlabs.io/docs/api-reference/voices/pvc/create"},"s925":{"title":"Voice changer overview","url":"https://elevenlabs.io/docs/overview/capabilities/voice-changer"},"s926":{"title":"Ollama v0.40.15 OpenAI request types","url":"https://github.com/ollama/ollama/blob/v0.40.15/openai/openai.go"},"s927":{"title":"Ollama vision","url":"https://docs.ollama.com/capabilities/vision"},"s928":{"title":"Ollama Anthropic compatibility","url":"https://docs.ollama.com/api/anthropic-compatibility"},"s929":{"title":"Ollama thinking","url":"https://docs.ollama.com/capabilities/thinking"},"s930":{"title":"Ollama tool calling","url":"https://docs.ollama.com/capabilities/tool-calling"},"s931":{"title":"Ollama Codex integration","url":"https://docs.ollama.com/integrations/codex"},"s932":{"title":"Ollama v0.40.15 Responses types","url":"https://github.com/ollama/ollama/blob/v0.40.15/openai/responses.go"},"s933":{"title":"Ollama v0.40.15 Anthropic web search middleware","url":"https://github.com/ollama/ollama/blob/v0.40.15/middleware/anthropic.go"},"s934":{"title":"Ollama usage metrics","url":"https://docs.ollama.com/api/usage"},"s935":{"title":"Ollama v0.40.15 registered routes","url":"https://github.com/ollama/ollama/blob/v0.40.15/server/routes.go"},"s936":{"title":"Ollama v0.40.15 Responses compaction","url":"https://github.com/ollama/ollama/blob/v0.40.15/server/responses_compact.go"},"s937":{"title":"Ollama speech","url":"https://docs.ollama.com/capabilities/speech"},"s938":{"title":"Ollama embeddings","url":"https://docs.ollama.com/capabilities/embeddings"},"s939":{"title":"Ollama v0.40.4 release notes","url":"https://github.com/ollama/ollama/releases/tag/v0.40.4"},"s940":{"title":"Ollama generate embeddings","url":"https://docs.ollama.com/api/embed"},"s941":{"title":"Ollama decision models","url":"https://docs.ollama.com/capabilities/decision"},"s942":{"title":"Ollama System One API","url":"https://docs.ollama.com/api/systemone"},"s943":{"title":"Ollama streaming","url":"https://docs.ollama.com/capabilities/streaming"},"s944":{"title":"Ollama chat API","url":"https://docs.ollama.com/api/chat"},"s945":{"title":"Ollama Cloud OpenAI-compatible model list","url":"https://ollama.com/v1/models"},"s946":{"title":"Ollama API introduction","url":"https://docs.ollama.com/api/introduction"},"s947":{"title":"Ollama OpenAPI specification","url":"https://docs.ollama.com/openapi.yaml"},"s948":{"title":"Ollama cloud usage API","url":"https://docs.ollama.com/api/cloud-usage"},"s949":{"title":"Ollama pricing","url":"https://ollama.com/pricing"},"s950":{"title":"Ollama Clef model","url":"https://ollama.com/library/clef"},"s951":{"title":"vLLM structured outputs","url":"https://docs.vllm.ai/en/latest/features/structured_outputs/"},"s952":{"title":"GPUStack vision language models","url":"https://docs.gpustack.ai/latest/using-models/using-vision-language-models/"},"s953":{"title":"vLLM reasoning outputs","url":"https://docs.vllm.ai/en/latest/features/reasoning_outputs/"},"s954":{"title":"GPUStack v2.3.0 release notes","url":"https://github.com/gpustack/gpustack/releases/tag/v2.3.0"},"s955":{"title":"GPUStack tool calling tutorial","url":"https://docs.gpustack.ai/latest/tutorials/inference-with-tool-calling/"},"s956":{"title":"vLLM tool calling","url":"https://docs.vllm.ai/en/latest/features/tool_calling/"},"s957":{"title":"vLLM v0.32.0 Responses tool dispatch source","url":"https://raw.githubusercontent.com/vllm-project/vllm/v0.32.0/vllm/entrypoints/openai/responses/context.py"},"s958":{"title":"vLLM v0.32.0 MCP tool server source","url":"https://raw.githubusercontent.com/vllm-project/vllm/v0.32.0/vllm/entrypoints/mcp/tool_server.py"},"s959":{"title":"vLLM v0.32.0 Responses serving source","url":"https://raw.githubusercontent.com/vllm-project/vllm/v0.32.0/vllm/entrypoints/openai/responses/serving.py"},"s960":{"title":"GPUStack usage metrics","url":"https://docs.gpustack.ai/latest/user-guide/usage/"},"s961":{"title":"GPUStack cache service management","url":"https://docs.gpustack.ai/latest/user-guide/cache-service-management/"},"s962":{"title":"vLLM prefix caching","url":"https://docs.vllm.ai/en/latest/design/prefix_caching/"},"s963":{"title":"vLLM v0.32.0 tokenizer request schemas","url":"https://raw.githubusercontent.com/vllm-project/vllm/v0.32.0/vllm/entrypoints/serve/tokenize/protocol.py"},"s964":{"title":"vLLM-Omni API server endpoints","url":"https://raw.githubusercontent.com/vllm-project/vllm-omni/v0.32.0/docs/serving/README.md"},"s965":{"title":"vLLM-Omni Videos API","url":"https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/videos_api/"},"s966":{"title":"vLLM-Omni Speech API","url":"https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/speech_api/"},"s967":{"title":"GPUStack embedding models","url":"https://docs.gpustack.ai/latest/using-models/using-embedding-models/"},"s968":{"title":"GPUStack reranker models","url":"https://docs.gpustack.ai/latest/using-models/using-reranker-models/"},"s969":{"title":"vLLM classification usages","url":"https://docs.vllm.ai/en/latest/models/pooling_models/classify/"},"s970":{"title":"GPUStack model deployment management","url":"https://docs.gpustack.ai/latest/user-guide/model-deployment-management/"},"s971":{"title":"GPUStack LoRA adapters","url":"https://docs.gpustack.ai/latest/tutorials/serving-with-lora-adapters/"},"s972":{"title":"vLLM-Omni Audio Generate API","url":"https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/audio_generate_api/"},"s973":{"title":"GPUStack model provider management","url":"https://docs.gpustack.ai/latest/user-guide/model-provider-management/"},"s974":{"title":"GPUStack Claude Code integration","url":"https://docs.gpustack.ai/latest/integrations/integrate-with-claude-code/"},"s975":{"title":"GPUStack model route management","url":"https://docs.gpustack.ai/latest/user-guide/model-route-management/"},"s976":{"title":"GPUStack managing model routes","url":"https://docs.gpustack.ai/latest/tutorials/managing-model-routes/"},"s977":{"title":"Hetzner Inference API","url":"https://docs.hetzner.com/general/company-and-policy/experiments/inference/"},"s978":{"title":"Hetzner Experiments platform overview","url":"https://docs.hetzner.com/general/company-and-policy/experiments/experiments-platform/"},"s979":{"title":"Choosing an API","url":"https://dev.meta.ai/docs/protocols"},"s980":{"title":"Responses API","url":"https://dev.meta.ai/docs/protocols/responses"},"s981":{"title":"Models","url":"https://dev.meta.ai/docs/models"},"s982":{"title":"Structured output","url":"https://dev.meta.ai/docs/structured-output"},"s983":{"title":"Image understanding","url":"https://dev.meta.ai/docs/image-understanding"},"s984":{"title":"File handling","url":"https://dev.meta.ai/docs/file-handling"},"s985":{"title":"Video and audio understanding","url":"https://dev.meta.ai/docs/video-understanding"},"s986":{"title":"Responses schemas","url":"https://dev.meta.ai/docs/api-reference/responses/schemas"},"s987":{"title":"Reasoning","url":"https://dev.meta.ai/docs/reasoning"},"s988":{"title":"Search grounding","url":"https://dev.meta.ai/docs/search-grounding"},"s989":{"title":"Tool calling","url":"https://dev.meta.ai/docs/tool-calling"},"s990":{"title":"Pricing and rate limits","url":"https://dev.meta.ai/docs/pricing-rate-limits"},"s991":{"title":"Image generation with Muse Image","url":"https://dev.meta.ai/docs/image-generation"},"s992":{"title":"API reference","url":"https://dev.meta.ai/docs/api-reference"},"s993":{"title":"Meta Model API documentation index","url":"https://dev.meta.ai/llms.txt"},"s994":{"title":"Build a computer-use agent","url":"https://dev.meta.ai/docs/computer-use"},"s995":{"title":"Prompt caching","url":"https://dev.meta.ai/docs/prompt-caching"},"s996":{"title":"Token counting","url":"https://dev.meta.ai/docs/token-counting"},"s997":{"title":"Responses API reference","url":"https://dev.meta.ai/docs/api-reference/responses"},"s998":{"title":"Generate an image","url":"https://dev.meta.ai/docs/api-reference/images/create-image"},"s999":{"title":"Edit an image","url":"https://dev.meta.ai/docs/api-reference/images/edit-image"},"s1000":{"title":"Speech to text","url":"https://dev.meta.ai/docs/speech-to-text"},"s1001":{"title":"Transcribe a recording","url":"https://dev.meta.ai/docs/api-reference/voice/transcribe"},"s1002":{"title":"Transcribe in realtime","url":"https://dev.meta.ai/docs/api-reference/voice/realtime"},"s1003":{"title":"Files API reference","url":"https://dev.meta.ai/docs/api-reference/files"},"s1004":{"title":"Files schemas","url":"https://dev.meta.ai/docs/api-reference/files/schemas"},"s1005":{"title":"Chat completions schemas","url":"https://dev.meta.ai/docs/api-reference/chat-completions/schemas"},"s1006":{"title":"Meta Model API overview","url":"https://dev.meta.ai/docs/overview"},"s1007":{"title":"Tool search","url":"https://dev.meta.ai/docs/tool-search"},"s1008":{"title":"MiniMax Anthropic SDK compatibility","url":"https://platform.minimax.io/docs/api-reference/text-anthropic-api"},"s1009":{"title":"MiniMax Create Message","url":"https://platform.minimax.io/docs/api-reference/text-chat-anthropic"},"s1010":{"title":"MiniMax Create Response","url":"https://platform.minimax.io/docs/api-reference/responses-create"},"s1011":{"title":"MiniMax Text Generation","url":"https://platform.minimax.io/docs/api-reference/text-post"},"s1012":{"title":"MiniMax model list","url":"https://platform.minimax.io/docs/faq/models"},"s1013":{"title":"MiniMax Upload File","url":"https://platform.minimax.io/docs/api-reference/file-management-upload"},"s1014":{"title":"MiniMax Create Chat Completion","url":"https://platform.minimax.io/docs/api-reference/text-chat-openai"},"s1015":{"title":"MiniMax Server Tools","url":"https://platform.minimax.io/docs/guides/server-tools"},"s1016":{"title":"MiniMax API overview","url":"https://platform.minimax.io/docs/api-reference/api-overview"},"s1017":{"title":"MiniMax Prompt Caching","url":"https://platform.minimax.io/docs/api-reference/text-prompt-caching"},"s1018":{"title":"MiniMax Estimate Input Tokens","url":"https://platform.minimax.io/docs/api-reference/responses-input-tokens"},"s1019":{"title":"MiniMax Image Generation guide","url":"https://platform.minimax.io/docs/guides/image-generation"},"s1020":{"title":"MiniMax Text to Image Generation","url":"https://platform.minimax.io/docs/api-reference/image-generation-t2i"},"s1021":{"title":"MiniMax Image-to-Image Generation","url":"https://platform.minimax.io/docs/api-reference/image-generation-i2i"},"s1022":{"title":"MiniMax Create Video Generation Task","url":"https://platform.minimax.io/docs/api-reference/video-generation-v2-create"},"s1023":{"title":"MiniMax Video Generation guide","url":"https://platform.minimax.io/docs/guides/video-generation"},"s1024":{"title":"MiniMax Synchronous Text to Speech","url":"https://platform.minimax.io/docs/api-reference/speech-t2a-http"},"s1025":{"title":"MiniMax Speech to Text API","url":"https://platform.minimax.io/docs/api-reference/speech-to-text"},"s1026":{"title":"MiniMax Speech-to-Text guide","url":"https://platform.minimax.io/docs/guides/speech-to-text"},"s1027":{"title":"MiniMax documentation index","url":"https://platform.minimax.io/docs/llms.txt"},"s1028":{"title":"MiniMax Retrieve Content","url":"https://platform.minimax.io/docs/api-reference/file-management-retrieve-content"},"s1029":{"title":"MiniMax List Files","url":"https://platform.minimax.io/docs/api-reference/file-management-list"},"s1030":{"title":"MiniMax Delete File","url":"https://platform.minimax.io/docs/api-reference/file-management-delete"},"s1031":{"title":"MiniMax self-hosting guide","url":"https://platform.minimax.io/docs/guides/local-deploy-h3"},"s1032":{"title":"MiniMax Music Generation","url":"https://platform.minimax.io/docs/api-reference/music-generation"},"s1033":{"title":"MiniMax WebSocket Text to Speech","url":"https://platform.minimax.io/docs/api-reference/speech-t2a-websocket"},"s1034":{"title":"MiniMax Bidirectional Streaming Text to Speech","url":"https://platform.minimax.io/docs/api-reference/speech-t2a-websocket-bidi"},"s1035":{"title":"MiniMax M Plan MCP guide","url":"https://platform.minimax.io/docs/guides/token-plan-mcp-guide"},"s1036":{"title":"MiniMax Create Video Regeneration Task","url":"https://platform.minimax.io/docs/api-reference/video-generation-v2-regeneration"},"s1037":{"title":"MiniMax Voice Clone API","url":"https://platform.minimax.io/docs/api-reference/voice-cloning-clone"},"s1038":{"title":"MiniMax Voice Clone guide","url":"https://platform.minimax.io/docs/guides/speech-voice-clone"},"s1039":{"title":"Chat Completions API","url":"https://platform.kimi.ai/docs/api/chat"},"s1040":{"title":"Model List","url":"https://platform.kimi.ai/docs/models"},"s1041":{"title":"Streaming","url":"https://platform.kimi.ai/docs/guide/utilize-the-streaming-output-feature-of-kimi-api"},"s1042":{"title":"Use response_format","url":"https://platform.kimi.ai/docs/guide/response_format"},"s1043":{"title":"Configure Kimi Vision Models","url":"https://platform.kimi.ai/docs/guide/use-kimi-vision-model"},"s1044":{"title":"Ask questions about PDF content","url":"https://platform.kimi.ai/docs/guide/ask-questions-about-pdf-content"},"s1045":{"title":"Upload File","url":"https://platform.kimi.ai/docs/api/files-upload"},"s1046":{"title":"API Overview","url":"https://platform.kimi.ai/docs/api/overview"},"s1047":{"title":"Platform Changelog","url":"https://platform.kimi.ai/docs/platform-changelog"},"s1048":{"title":"Thinking Models","url":"https://platform.kimi.ai/docs/guide/use-thinking-models"},"s1049":{"title":"Model Parameter Reference","url":"https://platform.kimi.ai/docs/api/models-overview"},"s1050":{"title":"Reasoning Effort","url":"https://platform.kimi.ai/docs/guide/use-reasoning-effort"},"s1051":{"title":"Responses API","url":"https://platform.kimi.ai/docs/api/responses"},"s1052":{"title":"Use Kimi API for Tool Calls","url":"https://platform.kimi.ai/docs/guide/use-kimi-api-to-complete-tool-calls"},"s1053":{"title":"Tool Choice","url":"https://platform.kimi.ai/docs/guide/use-tool-choice"},"s1054":{"title":"How to Use Official Tools","url":"https://platform.kimi.ai/docs/guide/use-official-tools"},"s1055":{"title":"Configure ModelScope MCP Server in Playground","url":"https://platform.kimi.ai/docs/guide/configure-the-modelscope-mcp-server"},"s1056":{"title":"Best Practices for context caching","url":"https://platform.kimi.ai/docs/guide/context-caching"},"s1057":{"title":"Messages API","url":"https://platform.kimi.ai/docs/api/messages"},"s1058":{"title":"Multi-turn Chat","url":"https://platform.kimi.ai/docs/guide/engage-in-multi-turn-conversations-using-kimi-api"},"s1059":{"title":"Estimate Tokens","url":"https://platform.kimi.ai/docs/api/estimate"},"s1060":{"title":"Get File Content","url":"https://platform.kimi.ai/docs/api/files-content"},"s1061":{"title":"Using Batch API","url":"https://platform.kimi.ai/docs/guide/use-batch-api"},"s1062":{"title":"Create Batch","url":"https://platform.kimi.ai/docs/api/batch-create"},"s1063":{"title":"Kimi K3.1 Pro","url":"https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart"},"s1064":{"title":"List Files","url":"https://platform.kimi.ai/docs/api/files-list"},"s1065":{"title":"Delete File","url":"https://platform.kimi.ai/docs/api/files-delete"},"s1066":{"title":"JSON Mode","url":"https://platform.kimi.ai/docs/guide/use-json-mode-feature-of-kimi-api"},"s1067":{"title":"Partial Mode","url":"https://platform.kimi.ai/docs/guide/use-partial-mode-feature-of-kimi-api"},"s1068":{"title":"Web Search Basic","url":"https://platform.kimi.ai/docs/api/tools-search"},"s1069":{"title":"Web Search Pro","url":"https://platform.kimi.ai/docs/api/tools-search-pro"},"s1070":{"title":"Best Practices for Web Search","url":"https://platform.kimi.ai/docs/guide/best-practices-for-web-search"},"s1071":{"title":"Kimi K3 Tool Calling Best Practices","url":"https://platform.kimi.ai/docs/guide/kimi-k3-tool-calling-best-practice"},"s1072":{"title":"Dynamically Loaded Tools","url":"https://platform.kimi.ai/docs/guide/use-dynamic-tool-loading"},"s1073":{"title":"URL Fetch","url":"https://platform.kimi.ai/docs/api/tools-fetch"},"s1074":{"title":"TypeSafe API reference: System One","url":"https://docs.typesafe.ai/api/system-one"},"s1075":{"title":"TypeSafe API reference: List models","url":"https://docs.typesafe.ai/api/list-models"},"s1076":{"title":"TypeSafe documentation index","url":"https://docs.typesafe.ai/llms.txt"},"s1077":{"title":"TypeSafe OpenAPI specification","url":"https://api.typesafe.ai/openapi.json"},"s1078":{"title":"TypeSafe Jev with coding agents","url":"https://docs.typesafe.ai/introduction/coding-agents"},"s1079":{"title":"TypeSafe Model jaggedness","url":"https://docs.typesafe.ai/model-jaggedness"},"s1080":{"title":"TypeSafe Images","url":"https://docs.typesafe.ai/images"},"s1081":{"title":"TypeSafe Models","url":"https://docs.typesafe.ai/models"},"s1082":{"title":"TypeSafe Double-checking citations cookbook","url":"https://docs.typesafe.ai/cookbooks/citation_check"},"s1083":{"title":"TypeSafe Function calling cookbook","url":"https://docs.typesafe.ai/cookbooks/function_calling"},"s1084":{"title":"TypeSafe Guardrails for LLMs cookbook","url":"https://docs.typesafe.ai/cookbooks/llm_guardrails"},"s1085":{"title":"TypeSafe Reading a scanned vaccination record cookbook","url":"https://docs.typesafe.ai/cookbooks/reading_scanned_documents"},"s1086":{"title":"TypeSafe Box","url":"https://docs.typesafe.ai/primitives/box"},"s1087":{"title":"TypeSafe Re-ranking cookbook","url":"https://docs.typesafe.ai/cookbooks/rerank_typesafe"},"s1088":{"title":"TypeSafe Parallel questions cookbook","url":"https://docs.typesafe.ai/cookbooks/parallel_questions"},"s1089":{"title":"TypeSafe Agent skill","url":"https://docs.typesafe.ai/agent-skill"},"s1090":{"title":"TypeSafe Python SDK asynchronous client","url":"https://docs.typesafe.ai/sdk/python/api/clients/async"},"s1091":{"title":"TypeSafe Choice","url":"https://docs.typesafe.ai/primitives/choice"},"s1092":{"title":"TypeSafe Classification using confidence cookbook","url":"https://docs.typesafe.ai/cookbooks/classification_using_confidence"},"s1093":{"title":"TypeSafe Line-by-line search cookbook","url":"https://docs.typesafe.ai/cookbooks/semantic_find"},"s1094":{"title":"TypeSafe Model changelog","url":"https://docs.typesafe.ai/changelog"},"s1095":{"title":"Z.ai Chat Completion","url":"https://docs.z.ai/api-reference/llm/chat-completion"},"s1096":{"title":"Z.ai Streaming Messages","url":"https://docs.z.ai/guides/capabilities/streaming"},"s1097":{"title":"Z.ai Structured Output","url":"https://docs.z.ai/guides/capabilities/struct-output"},"s1098":{"title":"Z.ai GLM-5.3-Flash/FlashX","url":"https://docs.z.ai/guides/vlm/glm-5.3-flash"},"s1099":{"title":"Z.ai Deep Thinking","url":"https://docs.z.ai/guides/capabilities/thinking"},"s1100":{"title":"Z.ai Thinking Mode","url":"https://docs.z.ai/guides/capabilities/thinking-mode"},"s1101":{"title":"Z.ai Web Search guide","url":"https://docs.z.ai/guides/tools/web-search"},"s1102":{"title":"Z.ai Function Calling","url":"https://docs.z.ai/guides/capabilities/function-calling"},"s1103":{"title":"Z.ai Pricing","url":"https://docs.z.ai/guides/overview/pricing"},"s1104":{"title":"Z.ai Official Python SDK","url":"https://docs.z.ai/guides/develop/python/introduction"},"s1105":{"title":"Z.ai documentation index","url":"https://docs.z.ai/llms.txt"},"s1106":{"title":"Z.ai OpenAPI specification","url":"https://docs.z.ai/openapi.json"},"s1107":{"title":"Z.ai Web Reader","url":"https://docs.z.ai/api-reference/tools/web-reader"},"s1108":{"title":"Z.ai Web Search MCP Server","url":"https://docs.z.ai/devpack/mcp/search-mcp-server"},"s1109":{"title":"Z.ai Context Caching","url":"https://docs.z.ai/guides/capabilities/cache"},"s1110":{"title":"Z.ai FAQ","url":"https://docs.z.ai/help/faq"},"s1111":{"title":"Z.ai Tokenizer","url":"https://docs.z.ai/api-reference/tools/tokenizer"},"s1112":{"title":"Z.ai Generate Image","url":"https://docs.z.ai/api-reference/image/generate-image"},"s1113":{"title":"Z.ai GLM-Image","url":"https://docs.z.ai/guides/image/glm-image"},"s1114":{"title":"Z.ai Generate Video","url":"https://docs.z.ai/api-reference/video/generate-video"},"s1115":{"title":"Z.ai Audio Transcriptions","url":"https://docs.z.ai/api-reference/audio/audio-transcriptions"},"s1116":{"title":"Z.ai GLM-ASR-2512","url":"https://docs.z.ai/guides/audio/glm-asr-2512"},"s1117":{"title":"Z.ai Layout Parsing","url":"https://docs.z.ai/api-reference/tools/layout-parsing"},"s1118":{"title":"Z.ai GLM-OCR","url":"https://docs.z.ai/guides/vlm/glm-ocr"},"s1119":{"title":"Z.ai File Upload","url":"https://docs.z.ai/api-reference/agents/file-upload"},"s1120":{"title":"Z.ai Overview","url":"https://docs.z.ai/guides/overview/overview"},"s1121":{"title":"Z.ai Agent Chat","url":"https://docs.z.ai/api-reference/agents/agent"},"s1122":{"title":"Z.ai Agent Conversation History","url":"https://docs.z.ai/api-reference/agents/agent-conversation"},"s1123":{"title":"Z.ai API Key Management Guide","url":"https://docs.z.ai/devpack/notice/api-key-management"},"s1124":{"title":"Z.ai GLM Coding Plan Quick Start","url":"https://docs.z.ai/devpack/quick-start"},"s1125":{"title":"Z.ai Web Search API","url":"https://docs.z.ai/api-reference/tools/web-search"}},"working_tree":false,"providers":[{"id":"openai","name":"OpenAI","notes":"The audit covers the public OpenAI inference APIs, built-in tools and their supporting resources: Responses, Chat Completions, Decisions, media, Files and Batch. The Agents API, GPT-Live, Multi-agent, organization administration and evals have no shared rows or are not exhaustively audited.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"Chat runs on the Responses API by default, and on Chat Completions for audio and search models or when you select that protocol.","sources":["s439","s440"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality openai/gpt-5.6-luna can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_openai_gpt-5_6-luna_can_have_a_basic_conversation.yml","notes":"Basic exchange through Responses."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Both Responses and Chat Completions stream through server-sent events into chunks and a final Message.","sources":["s441","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses openai/gpt-5.6-luna supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_openai_gpt-5_6-luna_supports_streaming_responses.yml","notes":"Streams through Responses."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema renders text.format with a JSON Schema on Responses and response_format on Chat Completions. Availability depends on the model.","sources":["s442","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"schema_format"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat#with_schema with openai/gpt-5.6-luna accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_openai_gpt-5_6-luna_accepts_a_json_schema_and_returns_structured_output.yml","notes":"Schema request and parsed result."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as input_image on Responses and image_url on Chat Completions, by URL, data URI or uploaded file ID, with low, high and original detail. Original detail applies to supported models.","sources":["s443"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/media.rb","method":"format_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_image"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_image"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models openai/gpt-5.6-luna can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_openai_gpt-5_6-luna_can_understand_local_images.yml","notes":"Local image as input_image."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDFs and other documents render as Responses input_file through file_data, file_url or file_id. Chat Completions accepts PDFs only, and not by external URL. Large files upload automatically to the Files API.","sources":["s444"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/media.rb","method":"format_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_document"},{"version":"v2","path":"lib/ruby_llm/providers/openai/responses.rb","method":"document_urls?"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_document"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models openai/gpt-5.6-luna understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_openai_gpt-5_6-luna_understands_pdfs.yml","notes":"Inline PDF as input_file."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat document URL models openai/gpt-5.6-luna reads PDFs the provider fetches from their URL","cassette":"spec/fixtures/vcr_cassettes/chat_document_url_models_openai_gpt-5_6-luna_reads_pdfs_the_provider_fetches_from_their_url.yml","notes":"PDF fetched by OpenAI from file_url."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Audio attachments render as input_audio on Chat Completions, which RubyLLM selects for audio model IDs. The Responses protocol rejects audio attachments. OpenAI retires the gpt-audio and gpt-4o-audio families on January 20, 2027 in favor of GPT Audio 1.5.","sources":["s445","s446"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/media.rb","method":"format_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_audio"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-audio-mini","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models openai/gpt-audio-mini can understand audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_openai_gpt-audio-mini_can_understand_audio.yml","notes":"input_audio on Chat Completions."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The Responses input schema and the images and vision guide document no video content part. Sampling frames yourself is image input.","sources":["s439","s443"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Reasoning summaries, encrypted reasoning state and reasoning token counts come back on the Message, and replay across turns while store stays false. Models return summaries, not raw reasoning, and only when asked or configured.","sources":["s447","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_reasoning_summary, parse_reasoning_signature, render_reasoning_item"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.4","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking openai/gpt-5.4 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_openai_gpt-5_4_returns_thinking_when_available.yml","notes":"Reasoning state through Responses."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking sets reasoning.effort and a summary request on Responses, and reasoning_effort on Chat Completions. Accepted levels depend on the model. reasoning.mode for Pro mode goes through provider_options.","sources":["s447","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"resolve_effort"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/thinking.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.4","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking openai/gpt-5.4 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_openai_gpt-5_4_returns_thinking_when_available.yml","notes":"Effort sent on the request."}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"URL, file search and container file annotations become typed Citation objects with offsets that survive joined output parts, in sync and streamed replies. Chat Completions search models return url_citation annotations.","sources":["s148","s145"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"parse_completion_response"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_citations, parse_file_citation"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/citation.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5-search-api","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with openai/gpt-5-search-api returns url citations when web search is enabled","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_openai_gpt-5-search-api_returns_url_citations_when_web_search_is_enabled.yml","notes":"url_citation on Chat Completions."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Ruby tools render as flat function definitions on Responses and nested ones on Chat Completions. Call results replay from local history. GPT-6 Astra needs Responses for tool calling.","sources":["s448"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"tool_for, render_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_function_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling openai/gpt-5.6-luna can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_openai_gpt-5_6-luna_can_use_tools.yml","notes":"Function call and result through Responses."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps to auto, none, required and a named function.","sources":["s448","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control openai/gpt-5.6-luna respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_openai_gpt-5_6-luna_respects_specific_tool_choice.yml","notes":"Named function choice."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several function_call items in one response become several tool calls, indexed by call ID.","sources":["s448"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_function_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling openai/gpt-5.6-luna can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_openai_gpt-5_6-luna_can_use_parallel_tool_calls.yml","notes":"Several calls in one response."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls:) sets parallel_tool_calls on the request.","sources":["s448","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control openai/gpt-5.6-luna respects calls: :one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_openai_gpt-5_6-luna_respects_calls_one_for_sequential_execution.yml","notes":"parallel_tool_calls set to false."}]},"server_web_search":{"v1":"native","v2":"native","offered":true,"notes":"The web_search alias adds the Responses tool and asks for its sources. Message#web_sources lists them, ServerToolCall keeps each search, and billed searches reach usage. gpt-5-search-api remains a Chat Completions search route. The older search preview models shut down on July 23, 2026.","sources":["s148","s446"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_provider_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_web_sources, parse_server_tool_use"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"gpt-5.2","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with openai/gpt-5.2 searches and records the tool call items with the sources they found","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_openai_gpt-5_2_searches_and_records_the_tool_call_items_with_the_sources_they_found.yml","notes":"Search items and sources."},{"status":"passed","recorded_on":"2026-11-05","model":"gpt-5.2","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with openai/gpt-5.2 streams searches and counts the ones it bills","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_openai_gpt-5_2_streams_searches_and_counts_the_ones_it_bills.yml","notes":"Streamed search with billed count."}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"OpenAI has no separate fetch tool. Reasoning models open pages and search within them as open_page and find_in_page actions of web_search, which RubyLLM keeps on ServerToolCall and lists through Message#web_sources.","sources":["s148"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_web_sources, parse_server_tool_items"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/web_source_spec.rb","example":"RubyLLM::WebSource Message#web_sources lists what a openai search found and opened, each address once","notes":"Parses a stubbed open_page action. The committed OpenAI search recordings contain search actions only."}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"The code_execution and code_interpreter aliases send code_interpreter with an automatic container. Container options merge in as keywords. Calls and outputs stay on ServerToolCall and replay from raw items.","sources":["s449"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat request rendering renders OpenAI Responses aliases","notes":"Request rendering only."}]},"server_file_search":{"v1":"missing","v2":"native","offered":true,"notes":"The file_search alias takes vector_store_ids as keywords, and file citations come back typed. Stores are created outside RubyLLM.","sources":["s145"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_file_citation, parse_server_tool_items"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"The mcp alias connects a remote server with url and name keywords. Approval requests appear as remote tool calls you approve or deny, then resume through mcp_approval_response items. Completed calls stay on ServerToolCall. Replay is stateless and streaming works.","sources":["s337","s3"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/approvals.rb","method":"parse_tool_approvals, render_tool_approval_response"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb"},{"version":"v2","path":"lib/ruby_llm/tool_call.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/chat_methods.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/tool_call.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openai executes a remote read-only tool after approval using stateless continuation","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openai_executes_a_remote_read-only_tool_after_approval_using_stateless_continuation.yml","notes":"Approval, execution and stateless continuation."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openai denies a remote call without executing it","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openai_denies_a_remote_call_without_executing_it.yml","notes":"Denial path."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openai streams a remote approval and its completed result","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openai_streams_a_remote_approval_and_its_completed_result.yml","notes":"Streamed approval and continuation."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Caching is automatic on supported models. Cached reads and cache writes reach the Tokens buckets. with_caching(key:) sets prompt_cache_key.","sources":["s4"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params, parse_usage"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"apply_prompt_cache_params"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.2","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat prompt cache round-trip openai/gpt-5.2 reuses the prompt cache with a shared key","cassette":"spec/fixtures/vcr_cassettes/chat_prompt_cache_round-trip_openai_gpt-5_2_reuses_the_prompt_cache_with_a_shared_key.yml","notes":"Second request reports cached tokens."}]},"cache_boundaries":{"v1":"passthrough","v2":"native","offered":true,"notes":"cache_until_here renders prompt_cache_breakpoint on a content part and moves system messages into input items, since top-level instructions cannot hold one. Explicit breakpoints need GPT-5.6 or later.","sources":["s4"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/media.rb","method":"format_content"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"inject_cache_breakpoint, render_instructions"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat#with_caching renders explicit breakpoints for OpenAI cache boundaries","notes":"Request rendering only."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The prompt caching guide documents automatic caching, breakpoints and lifetimes. The reference lists no cache resource to create, read or delete.","sources":["s4","s450"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"compaction":{"v1":"missing","v2":"native","offered":true,"notes":"with_compaction sends context_management with a compaction entry and an optional compact_threshold through at:. The returned compaction item replays on later turns. The truncation request field is recorded under Context editing.","sources":["s351","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"apply_compaction, parse_server_tool_items"},{"version":"v2","path":"lib/ruby_llm/chat.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_compaction_spec.rb","example":"RubyLLM::Chat provider mapping maps to a compact_threshold entry on the OpenAI Responses API","notes":"Request rendering. The live recording covers the standalone endpoint, under Manual context compaction."}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"count_tokens posts to the Responses input_tokens endpoint and counts text, images, files, function tools, schemas and deferred tools. Hosted provider tools and provider_options are not part of the count request.","sources":["s306","s307"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/token_counting.rb","method":"render_count_tokens_payload, COUNT_TOKENS_KEYS"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with openai/gpt-5.6-luna counts a staged message without mutating the chat","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_openai_gpt-5_6-luna_counts_a_staged_message_without_mutating_the_chat.yml","notes":"Staged message count."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with openai/gpt-5.6-luna counts instructions, tools and structured output","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_openai_gpt-5_6-luna_counts_instructions_tools_and_structured_output.yml","notes":"Instructions, tools and schema."}]},"image_generation":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.paint posts to images/generations and returns one or several images. Quality and output format go through provider_options. SVG and animated GIF output apply to GPT Image 2.6 models.","sources":["s451","s452"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"render_image_payload, parse_image_responses"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-image-2.6-flare","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality openai/gpt-image-2.6-flare can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_openai_gpt-image-2_6-flare_can_paint_images.yml","notes":"Single image."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-image-1.5","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality openai/gpt-image-1.5 paints several images in one request","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_openai_gpt-image-1_5_paints_several_images_in_one_request.yml","notes":"Multiple images."}]},"image_editing":{"v1":"native","v2":"native","offered":true,"notes":"paint with with: and mask: posts to images/edits, as JSON image references for GPT Image models and as multipart uploads otherwise.","sources":["s451","s452"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"images_url, render_edit_payload, json_image_references?"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/chat_completions/images_spec.rb","example":"RubyLLM::Protocols::ChatCompletions::Images#images_url targets the edits endpoint for a bare IO without reading it","notes":"Endpoint selection. No committed OpenAI edit recording."}]},"video_generation":{"v1":"deprecated","v2":"na","offered":false,"notes":"OpenAI shut down the Videos API and Sora 2 models on September 24, 2026 and names no replacement. Video generation is not available. The remaining guide and reference pages are kept for history.","sources":["s446","s6","s453"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak posts to audio/speech and returns binary audio. OpenAI shuts down the models on this endpoint (tts-1, tts-1-hd, gpt-4o-mini-tts) on January 6, 2027 and points to Realtime, which RubyLLM does not implement.","sources":["s147","s454","s446"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"render_speech_payload, parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-4o-mini-tts","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality openai/gpt-4o-mini-tts can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_openai_gpt-4o-mini-tts_can_speak.yml","notes":"Complete speech result."}]},"transcription":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts files to audio/transcriptions with language, prompt, temperature, timestamps and response format. gpt-transcribe takes languages and keywords through provider_options, and the language: keyword sends the singular field that model replaces. whisper-1 and the gpt-4o transcribe models shut down on February 26, 2027.","sources":["s143","s455"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"render_transcription_payload, parse_transcription_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"whisper-1","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openai/whisper-1 can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openai_whisper-1_can_transcribe_audio.yml","notes":"File transcription."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe streams a recorded file into TranscriptionChunk objects and a final Transcription. Live audio sessions belong to Realtime transcription, recorded under Realtime sessions.","sources":["s143","s456"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"stream_transcription, build_transcription_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-4o-transcribe","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openai/gpt-4o-transcribe streams text deltas and returns the final transcription","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openai_gpt-4o-transcribe_streams_text_deltas_and_returns_the_final_transcription.yml","notes":"Text deltas and final event."}]},"diarization":{"v1":"native","v2":"native","offered":true,"notes":"Diarization models default to diarized_json, return speaker segments and accept known speaker names and references. The documented model, gpt-4o-transcribe-diarize, shuts down on February 26, 2027.","sources":["s143","s446"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"default_response_format, encode_speaker_references"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-4o-transcribe-diarize","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openai/gpt-4o-transcribe-diarize can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openai_gpt-4o-transcribe-diarize_can_transcribe_audio.yml","notes":"Speaker segments."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-4o-transcribe-diarize","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openai/gpt-4o-transcribe-diarize streams segments labelled with speakers","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openai_gpt-4o-transcribe-diarize_streams_segments_labelled_with_speakers.yml","notes":"Streamed speaker segments."}]},"moderation":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.moderate posts text and images to moderations and returns typed results.","sources":["s457","s458"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/moderation.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb","method":"render_moderation_payload, parse_moderation_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"omni-moderation-latest","spec":"spec/ruby_llm/moderation_spec.rb","example":"RubyLLM::Moderation.moderate with OpenAI provider moderates content and returns a Moderation instance","cassette":"spec/fixtures/vcr_cassettes/moderation_moderate_with_openai_provider_moderates_content_and_returns_a_moderation_instance.yml","notes":"Text moderation."},{"status":"passed","recorded_on":"2026-10-11","model":"omni-moderation-latest","spec":"spec/ruby_llm/moderation_spec.rb","example":"RubyLLM::Moderation.moderate with OpenAI provider moderates an image attachment without text","cassette":"spec/fixtures/vcr_cassettes/moderation_moderate_with_openai_provider_moderates_an_image_attachment_without_text.yml","notes":"Image moderation."}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog document no OCR or document parsing endpoint. File inputs and high-detail images go through a chat model, which is not an OCR API.","sources":["s137","s450","s459","s444"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.embed posts one or several texts to embeddings and supports dimensions.","sources":["s460","s461"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"text-embedding-3-small","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality openai/text-embedding-3-small can handle a single text with custom dimensions","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_openai_text-embedding-3-small_can_handle_a_single_text_with_custom_dimensions.yml","notes":"Custom dimensions."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The embeddings reference accepts strings and token arrays only.","sources":["s461"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog document no rerank endpoint. File search ranking options apply inside that tool.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to the Files API with purpose and expires_in, and UploadedFile.find reads metadata. Uploaded IDs attach to chat. Multipart Uploads for larger files have no integration.","sources":["s462","s444"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/files.rb","method":"render_upload_fields, parse_file_response"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload, find"}],"verification":[{"status":"passed","recorded_on":"2026-10-25","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads openai uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_openai_uploads_a_pdf_through_the_files_api.yml","notes":"Upload through the Files API."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads openai/gpt-5.6-luna reuses an uploaded file in chat","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_openai_gpt-5_6-luna_reuses_an_uploaded_file_in_chat.yml","notes":"file_id in a chat request."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download reads files/{id}/content and returns a DownloadedFile.","sources":["s462"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"download, download_file_url"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/files.rb"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch uploads JSONL and creates batches for Responses or Chat Completions, then reads results back by custom_id, with a 50 percent cost multiplier.","sources":["s7","s463"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"batch_protocol_name_for, batch_cost_multiplier"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/batches.rb","method":"create_batch, batch_results"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/batches.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openai/gpt-5.6-luna answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_openai_gpt-5_6-luna_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Full batch lifecycle."}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.embed_later targets /v1/embeddings through the Batch API and checks record positions in the results.","sources":["s7"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"batch_protocol_name_for"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embedding_batches.rb","method":"parse_batch_result"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"text-embedding-3-small","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openai/text-embedding-3-small embeddings embeds staged texts and hydrates each request result","cassette":"spec/fixtures/vcr_cassettes/batch_with_openai_text-embedding-3-small_embeddings_embeds_staged_texts_and_hydrates_each_request_result.yml","notes":"Embedding batch lifecycle."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document an advisor tool..","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Perplexity Agent API, which is not an API of this provider. OpenAI's own Agents API is a separate hosted session product outside this row.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"agent_skills":{"v1":"missing","v2":"passthrough","offered":true,"notes":"A raw shell tool Hash passed to with_provider_tools can reference skills prepared elsewhere through environment.skills. RubyLLM has no Skills upload or version API. RubyLLM's own Skill class runs locally and does not touch this API.","sources":["s138","s464"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw, normalize_positional"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"async_research":{"v1":"missing","v2":"na","offered":false,"notes":"OpenAI shut down its deep research models on July 23, 2026 and documents no dedicated research API. Long reasoning runs use background mode, recorded under Background Responses jobs.","sources":["s446","s465","s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"background_responses":{"v1":"missing","v2":"partial","offered":true,"notes":"provider_options can set background: true, so a job can start. RubyLLM has no retrieve, poll or cancel path for the response, so the result stays out of reach and the returned Message holds no answer.","sources":["s140","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"browser_use":{"v1":"missing","v2":"partial","offered":true,"notes":"A Ruby function tool can drive your own browser through the code-execution pattern in OpenAI's guide. The built-in computer tool goes through a raw Hash, its computer_call is kept as a ServerToolCall, and RubyLLM has no way to return computer_call_output. The hosted browser belongs to the Agents API.","sources":["s141"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"classification":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.judge posts to the Decisions API, whose choice, predicate and score questions classify text and images against fixed options. RubyLLM selects this protocol for judgments. gpt-6-luna is the only supported model.","sources":["s466","s467"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/decisions.rb","method":"render_judgment_payload, parse_judgment_response"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"resolve_protocol"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-6-luna","spec":"spec/ruby_llm/providers/openai_judgments_spec.rb","example":"RubyLLM::Providers::OpenAI with the Decisions API judges all three question types through the compact DSL","cassette":"spec/fixtures/vcr_cassettes/providers_openai_with_the_decisions_api_judges_all_three_question_types_through_the_compact_dsl.yml","notes":"Predicate, choice and score answers."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-6-luna","spec":"spec/ruby_llm/providers/openai_judgments_spec.rb","example":"RubyLLM::Providers::OpenAI with the Decisions API judges an image without text input","cassette":"spec/fixtures/vcr_cassettes/providers_openai_with_the_decisions_api_judges_an_image_without_text_input.yml","notes":"Image input."}]},"computer_use":{"v1":"missing","v2":"partial","offered":true,"notes":"A raw computer tool Hash is forwarded and the model's computer_call is kept as a ServerToolCall, but RubyLLM cannot send computer_call_output, so the action loop stops there. A Ruby function tool works for the code-execution pattern in OpenAI's guide.","sources":["s141"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"context_editing":{"v1":"na","v2":"passthrough","offered":true,"notes":"The Responses truncation field set to auto drops the oldest items when the input exceeds the context window, reachable through provider_options. OpenAI documents no selective tool result clearing.","sources":["s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document contextual embeddings..","sources":["s137","s450","s459","s461"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"dedicated_transcription_api":{"v1":"native","v2":"native","offered":true,"notes":"audio/transcriptions is a dedicated file transcription operation. This row overlaps Transcription and is not a separate gain.","sources":["s143","s454"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog document no dubbing operation. Realtime translation is a live session.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Fast and Ultrafast processing are selected with service_tier, which with_provider_options forwards. RubyLLM has no dedicated switch and no WebSocket route, which OpenAI recommends for Ultrafast. Ultrafast covers GPT-6 Astra and GPT-6.1 Sol.","sources":["s144","s468"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"The Files API lists and deletes files. UploadedFile offers upload, find and download only, and raw options cannot reach those endpoints.","sources":["s462"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/uploaded_file.rb","method":"upload, find, download"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Vector stores have their own creation, file indexing and search endpoints. RubyLLM can use a prepared vector_store_id with file_search but cannot manage the store.","sources":["s145","s469"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The legacy Completions reference still lists suffix, only for gpt-3.5-turbo-instruct, which OpenAI shut down on September 28, 2026. No current model is documented for it.","sources":["s470","s471"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"fine_tuning":{"v1":"missing","v2":"deprecated","offered":true,"notes":"OpenAI is winding down fine-tuning. Eligible existing customers can create jobs until January 6, 2027. RubyLLM has no fine-tuning job API. Calling an already fine-tuned model works as ordinary chat until its base model shuts down.","sources":["s472","s446","s473"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google Maps grounding, which is not an API of this provider.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"guardrail_configuration":{"v1":"na","v2":"passthrough","offered":true,"notes":"The moderation request field on Responses sets a moderation model and a score or block policy for input and output. with_provider_options forwards it. RubyLLM has no named setting and does not type the returned scores. Agents SDK guardrails are a client library feature outside this row.","sources":["s439","s457"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"OpenAI documents stored responses, previous_response_id and a Conversations resource. These are provider-stored conversation state, outside RubyLLM's coverage. RubyLLM sends store: false and builds requests from the local transcript.","sources":["s440","s474"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"image_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"The Batch API accepts /v1/images/generations and /v1/images/edits. RubyLLM batches raise for payloads that are not chat or embeddings, so raw options cannot reach them.","sources":["s7"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"batch_protocol_name_for, batch_protocol_for_endpoint"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Interactions API, which is not an API of this provider.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"JSON object mode goes through with_provider_options as text.format on Responses or response_format on Chat Completions. with_schema is the named API for schemas.","sources":["s442","s439"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_request_options_spec.rb","example":"RubyLLM::Chat with params openai/gpt-5.6-luna supports response_format param","cassette":"spec/fixtures/vcr_cassettes/chat_with_params_openai_gpt-5_6-luna_supports_response_format_param.yml","notes":"json_object through provider options."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names Amazon Bedrock Mantle, which is not an API of this provider.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"manual_compaction":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#compact posts to responses/compact and keeps the returned envelope. It replaces only the context of later requests. Local messages and instructions stay. Pending tool calls block it.","sources":["s351","s475"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb","method":"compaction_url, render_compaction_payload, parse_compaction_response"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_conversation"},{"version":"v2","path":"lib/ruby_llm/providers/openai/responses.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with openai","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_openai.yml","notes":"Compaction followed by another turn."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document music generation..","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"nonchat_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"The Batch API also accepts moderations, image endpoints and legacy completions. RubyLLM registers Responses, Chat Completions and embedding batches only, and raises for other payloads.","sources":["s7"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"batch_protocol_name_for"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row names a Google Omni video endpoint, which is not an API of this provider.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The API serves OpenAI's own models through its own wire formats. OpenAI models on Amazon Bedrock belong to that provider's coverage.","sources":["s137"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs document no assistant prefill or prefix continuation. Predicted Outputs speed up a known output and do not constrain the response to a prefix.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"programmatic_tool_calling":{"v1":"missing","v2":"partial","offered":true,"notes":"A raw programmatic_tool_calling Hash and allowed_callers through tool provider_options reach the API, and program items stay raw on ServerToolCall. A program that calls a local function cannot resume, because ToolCall drops the caller field that function_call_output must return.","sources":["s476"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items, parse_function_calls, render_tool_items"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching maps key, mode and ttl to prompt_cache_key and prompt_cache_options, including the 12-hour Responses lifetime. Those lifetime options apply to GPT-5.6 and later. Earlier models take prompt_cache_retention, which needs with_provider_options because the retention: keyword renders as ttl.","sources":["s4"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params, build_prompt_cache_options, retention_ttl"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"build_prompt_cache_options"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document provider-managed model fallback on request..","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"OpenAI documents the Realtime API and GPT-Live for live conversations and live transcription. These are outside RubyLLM's coverage. Conversation state stays in the application.","sources":["s477","s478","s456"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document a whole-response cache.. Prompt caching is recorded separately.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"Responses is the default OpenAI chat protocol. Audio, realtime and search model IDs route to Chat Completions, and the protocol option selects either one.","sources":["s439","s479"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/openai/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality openai/gpt-5.6-luna can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_openai_gpt-5_6-luna_can_have_a_basic_conversation.yml","notes":"Request posted to responses."}]},"responses_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch targets /v1/responses by default, with a dedicated serializer and result parser.","sources":["s7"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb","method":"batch_protocol_name_for, batch_protocol_for_endpoint"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/batches.rb","method":"batch_endpoint, parse_batch_completion_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openai/gpt-5.6-luna answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_openai_gpt-5_6-luna_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Batch on /v1/responses."}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs document no standalone web search endpoint. Vector store search is recorded under Vector store management.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"server_apply_patch":{"v1":"missing","v2":"partial","offered":true,"notes":"A raw apply_patch Hash is forwarded and proposed patches stay readable on ServerToolCall, but RubyLLM cannot send apply_patch_call_output, so the model never sees the result.","sources":["s142"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"server_tool_image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"The image_generation alias adds the hosted tool to Responses. Tool options merge in as keywords, and the result stays as base64 on ServerToolCall rather than a typed Image.","sources":["s480"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"server_tool_result"}]},"server_tool_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_tools(defer: true) and deferred tool classes send defer_loading and add the tool_search tool. Namespaced and searched calls dispatch back to Ruby tools and stay loaded on later turns. OpenAI offers it on gpt-5.4 and later.","sources":["s146"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_tool_search, render_tools, namespace_searched_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"searched_tool_namespace?, portable_raw_content"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.4","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with openai/gpt-5.4 loads a deferred tool through tool search and calls it","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_openai_gpt-5_4_loads_a_deferred_tool_through_tool_search_and_calls_it.yml","notes":"tool_search items and namespaced call."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.4","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with openai/gpt-5.4 keeps using a loaded tool on the next turn","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_openai_gpt-5_4_keeps_using_a_loaded_tool_on_the_next_turn.yml","notes":"Loaded tool on a later turn."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.4","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with openai/gpt-5.4 loads a deferred tool while streaming","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_openai_gpt-5_4_loads_a_deferred_tool_while_streaming.yml","notes":"Streamed tool search."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document an X search tool..","sources":["s137","s450","s459","s3"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"shell_computer_tools":{"v1":"missing","v2":"partial","offered":true,"notes":"Hosted shell goes through a raw shell Hash, with its calls and outputs kept on ServerToolCall and replayed. Local shell mode cannot run, because RubyLLM has no way to send shell_call_output. The computer tool has the same limit; see Computer use.","sources":["s481","s141"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document sound effect generation..","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak yields SpeechChunk bytes as they arrive and returns the complete Speech. The endpoint's models shut down on January 6, 2027.","sources":["s147","s446"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"stream_speech"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"stream_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-4o-mini-tts","spec":"spec/ruby_llm/speech_streaming_spec.rb","example":"RubyLLM::Speech streams speech from openai","cassette":"spec/fixtures/vcr_cassettes/speech_streams_speech_from_openai.yml","notes":"Chunked audio."}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The docs describe no token budget for a whole task. max_tool_calls limits built-in tool calls in a response and return_token_budget limits returned web search content. Both are narrower tool settings that provider_options and tool options can reach.","sources":["s439","s148"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs document no text analysis API returning intents, topics or sentiment. Decisions answers questions you write and is recorded under Classification.","sources":["s137","s450","s459"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"The vector_stores endpoints create, list, update, delete and search stores and their files. RubyLLM has no integration. Hosted file_search can use a prepared store.","sources":["s469"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"},{"version":"v2","path":"lib/ruby_llm/uploaded_file.rb"}]},"video_batches":{"v1":"deprecated","v2":"na","offered":false,"notes":"OpenAI shut down the Videos API and Sora 2 models on September 24, 2026 and names no replacement. Batch video generation is not available. The remaining guide and reference pages are kept for history.","sources":["s446","s6","s453","s7"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"video_editing":{"v1":"deprecated","v2":"na","offered":false,"notes":"OpenAI shut down the Videos API and Sora 2 models on September 24, 2026 and names no replacement. Video editing is not available. The remaining guide and reference pages are kept for history.","sources":["s446","s6","s453"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"video_extension":{"v1":"deprecated","v2":"na","offered":false,"notes":"OpenAI shut down the Videos API and Sora 2 models on September 24, 2026 and names no replacement. Video extension is not available. The remaining guide and reference pages are kept for history.","sources":["s446","s6","s453"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"voice_cloning":{"v1":"missing","v2":"missing","offered":true,"notes":"OpenAI documents custom voices from a consent recording and reference sample, and prompt-generated voices for GPT-Live. RubyLLM has no voice or consent API. A voice created elsewhere can be used in speech through provider_options.","sources":["s482","s454"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"render_speech_payload"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs index, API reference index and changelog do not document voice conversion..","sources":["s137","s450","s459","s454"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited docs document no standalone web fetch endpoint. Page opening exists only as a web_search action.","sources":["s137","s450","s459","s148"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"}]},"websockets":{"v1":"missing","v2":"missing","offered":true,"notes":"Responses WebSocket mode runs over a persistent connection to /v1/responses and works with store: false. RubyLLM uses HTTP for Responses. Its WebSocket transport serves other providers' transcription.","sources":["s483"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openai.rb"},{"version":"v2","path":"lib/ruby_llm/transport/websocket_connection.rb"}]}},"gaps":[{"notes":"Background Responses jobs: provider_options can start a job, but RubyLLM has no retrieve, poll or cancel path for its result.","sources":["s140"]},{"notes":"Browser use, Computer use, Apply patch tool: Raw tool definitions are forwarded and their calls stay readable on ServerToolCall. RubyLLM cannot send computer_call_output or apply_patch_call_output, so the client loop stops. Ruby function tools work for the code-execution pattern.","sources":["s141","s142"]},{"notes":"Programmatic tool calling: The hosted tool goes through a raw Hash. A program that calls a local function cannot resume, because ToolCall drops the caller field.","sources":["s476"]},{"notes":"Shell and computer tools: Hosted shell goes through a raw Hash. Local shell mode has no way to send shell_call_output.","sources":["s481"]},{"notes":"File listing and deletion, File search store management, Vector store management: RubyLLM has no list or delete for files and no vector store API. Hosted file_search can use a store prepared elsewhere.","sources":["s462","s469"]},{"notes":"Image batches, Other batch operations: The Batch API also accepts image, moderation and legacy completion endpoints. RubyLLM registers Responses, Chat Completions and embedding batches only.","sources":["s7"]},{"notes":"Voice cloning: RubyLLM has no voice or consent API. A voice created elsewhere can be used in speech through provider_options.","sources":["s482"]},{"notes":"Responses over WebSocket: RubyLLM uses HTTP for Responses and has no WebSocket mode, which OpenAI recommends for Ultrafast.","sources":["s483","s468"]},{"notes":"Fast inference, JSON object mode, Guardrail configuration, Context editing, Agent skills: These work through provider_options or a raw tool Hash without a named RubyLLM setting.","sources":["s144","s439","s138"]}]},{"id":"anthropic","name":"Anthropic","notes":"The audit covers the direct Claude API: Messages, Message Batches, Token Counting, Models, Files and Skills, with Managed Agents recorded as hosted sessions. RubyLLM sends API version 2023-06-01. Built-in coverage is conditional on supported models, and organization and billing administration are outside scope.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"chat, ask and agents reach POST /v1/messages on the Messages API. Model restrictions apply.","sources":["s186","s484"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"render_payload, parse_completion_body, completion_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality anthropic/claude-haiku-4-5 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_anthropic_claude-haiku-4-5_can_have_a_basic_conversation.yml","notes":"Basic Messages API exchange with token usage."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"ask with a block streams server-sent events, including text, thinking, tool input, citation and compaction deltas.","sources":["s485"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/streaming.rb","method":"build_chunk, track_stream_blocks"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses anthropic/claude-haiku-4-5 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_anthropic_claude-haiku-4-5_supports_streaming_responses.yml","notes":"Streamed chunks assemble into the final message."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema renders output_config.format as json_schema. Anthropic rejects structured outputs combined with citations.","sources":["s187","s35"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_output_config"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat#with_schema with anthropic/claude-haiku-4-5 accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_anthropic_claude-haiku-4-5_accepts_a_json_schema_and_returns_structured_output.yml","notes":"Request carries output_config and the reply parses to a Hash."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as base64, URL or Files API file_id sources.","sources":["s486"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"render_image, render_provider_image"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models anthropic/claude-haiku-4-5 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_anthropic_claude-haiku-4-5_can_understand_local_images.yml","notes":"Local image sent as an attachment."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDF attachments render as document blocks from base64, a URL or a Files API file_id.","sources":["s487"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"render_pdf, pdf_source, render_provider_file"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models anthropic/claude-haiku-4-5 understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_anthropic_claude-haiku-4-5_understands_pdfs.yml","notes":"PDF sent as a document block."}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The Messages API content blocks have no audio input. Chat attachments cover images, PDFs and text. RubyLLM raises UnsupportedAttachmentError for this type.","sources":["s484","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"supported_attachment?"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The Messages API content blocks have no video input. Chat attachments cover images, PDFs and text. RubyLLM raises UnsupportedAttachmentError for this type.","sources":["s484","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"supported_attachment?"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Thinking text, signatures and redacted blocks parse into Thinking values, stream, and replay. Recent models omit the text unless display is summarized.","sources":["s488"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_thinking_content, parse_thinking_signature, render_thinking_blocks"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-sonnet-5","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat thinking display with anthropic/claude-sonnet-5 returns readable thinking with display summarized","cassette":"spec/fixtures/vcr_cassettes/chat_thinking_display_with_anthropic_claude-sonnet-5_returns_readable_thinking_with_display_summarized.yml","notes":"Adaptive thinking with xhigh effort returns summarized text and a signature."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking maps effort to output_config.effort, budget to budget_tokens, and display to thinking.display, and picks adaptive or budgeted mode from the registry. Sonnet 5.5 turns off with between_tools. Opus 5.5 and Fable 5.1 reject turning thinking off, and the API returns that error. Per-message effort changes are not exposed.","sources":["s488","s489"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_thinking_payload, thinking_mode, thinking_off_type"},{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"between_tools_off?"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/block_binding.rb","method":"apply"}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_citations enables citations on text, PDF and Files API documents, and tool results can return search result blocks. Replies carry normalized Citation values, including while streaming.","sources":["s35"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb","method":"format_content"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"enable_citations, render_text_document"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_text_and_citations, parse_citation"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"search_result_block"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with anthropic/claude-haiku-4-5 cites text documents in responses","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_anthropic_claude-haiku-4-5_cites_text_documents_in_responses.yml","notes":"Citation offsets map back to the response text."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Ruby tools render as custom tools with input_schema, and tool_use blocks run in the chat loop.","sources":["s490"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"render_tools, function_for, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling anthropic/claude-haiku-4-5 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_anthropic_claude-haiku-4-5_can_use_tools.yml","notes":"A tool call and its result complete a turn."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps to auto, any, none and tool. Opus 5.5, Sonnet 5.5 and Fable 5.1 reject any and tool with a 400 error, which RubyLLM surfaces.","sources":["s490","s491"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control anthropic/claude-haiku-4-5 respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_anthropic_claude-haiku-4-5_respects_specific_tool_choice.yml","notes":"Request forces the named tool."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool_use blocks in one response all become tool calls.","sources":["s492"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/anthropic/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"find_tool_uses, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling anthropic/claude-haiku-4-5 can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_anthropic_claude-haiku-4-5_can_use_parallel_tool_calls.yml","notes":"One response asks for two tools."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) sets disable_parallel_tool_use inside tool_choice.","sources":["s492"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control anthropic/claude-haiku-4-5 respects calls: :one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_anthropic_claude-haiku-4-5_respects_calls_one_for_sequential_execution.yml","notes":"Request carries disable_parallel_tool_use."}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends web_search_20260318 pinned to direct callers. Replies carry ServerToolCall values, citations, web_sources and search counts, and pause_turn continues automatically.","sources":["s493"],"code":[{"version":"v1","path":"lib/ruby_llm/provider.rb","method":"complete"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES, complete"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_server_tool_calls, parse_web_sources"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with anthropic/claude-haiku-4-5 searches, cites, and reports tool usage","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_anthropic_claude-haiku-4-5_searches_cites_and_reports_tool_usage.yml","notes":"Search results, citations, sources and search count."}]},"server_web_fetch":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_fetch) or :url_context sends web_fetch_20260318 pinned to direct callers. Organization admins can turn web fetch off, and the API then returns a 400 error. No recording covers this alias.","sources":["s494","s491"],"code":[{"version":"v1","path":"lib/ruby_llm/provider.rb","method":"complete"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"server_tool_block?"}]},"server_code_execution":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:code_execution) sends code_execution_20260521, and result blocks return as ServerToolCall values. Generated file IDs stay in the raw result; fetch them with RubyLLM.download. This is the hosted tool, separate from with_code_execution, which runs in your own sandbox.","sources":["s495"],"code":[{"version":"v1","path":"lib/ruby_llm/provider.rb","method":"complete"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"server_tool_block?, render_raw_assistant_message"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat code execution with anthropic/claude-haiku-4-5 runs code server-side and returns the result blocks","cassette":"spec/fixtures/vcr_cassettes/chat_code_execution_with_anthropic_claude-haiku-4-5_runs_code_server-side_and_returns_the_result_blocks.yml","notes":"Request carries code_execution_20260521."}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The server tool list has no hosted file search. Search result blocks hold results your application supplies, and tool search finds tools, not documents.","sources":["s33","s186"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES"}]},"server_mcp":{"v1":"passthrough","v2":"native","offered":true,"notes":"The mcp alias configures mcp_servers and a matching mcp_toolset with the mcp-client-2025-11-20 beta. Tool selection uses allowed_tools or default_config and configs. Anthropic runs tools without approval, so RubyLLM accepts only require_approval: never. Replies replay as returned. The newer mcp-client-2026-09-15 additions need raw headers.","sources":["s36","s491"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES, MCP_BETA, mcp_tool_filter"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP execution anthropic executes a read-only tool and replays its history","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_execution_anthropic_executes_a_read-only_tool_and_replays_its_history.yml","notes":"Read-only Microsoft Learn MCP search, typed remote results and a follow-up replay."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"with_caching sends top-level cache_control for automatic caching, and usage reports cache reads and writes by lifetime. Token minimums and prefix rules still apply.","sources":["s37"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/content.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_control"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/models.rb","method":"parse_cache_write_by_ttl"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat prompt cache round-trip anthropic/claude-haiku-4-5 writes then reads the prompt cache","cassette":"spec/fixtures/vcr_cassettes/chat_prompt_cache_round-trip_anthropic_claude-haiku-4-5_writes_then_reads_the_prompt_cache.yml","notes":"A second request reads what the first wrote."}]},"cache_boundaries":{"v1":"native","v2":"native","offered":true,"notes":"cache_until_here and with_instructions(cache_until_here: true) put cache_control on the last block of a message, with a per-boundary lifetime.","sources":["s37"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic/content.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"cache_boundary?, inject_cache_control"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Prompt caching uses cache_control on requests. The API inventory has no cache resource to create, update or delete.","sources":["s37","s186"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_control"}]},"compaction":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_compaction renders a compact_20260112 edit with at, instructions and pause_after, sends the compact-2026-01-12 beta, keeps compaction blocks in history, and sums usage.iterations so the summarization pass is billed. The API requires a threshold of at least 50,000 tokens.","sources":["s496","s38"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction, compaction_edit, server_tool_block?"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/models.rb","method":"aggregate_usage"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-sonnet-4-6","spec":"spec/ruby_llm/chat_compaction_spec.rb","example":"RubyLLM::Chat anthropic claude-sonnet-4-6 compacts a long conversation and bills the summarization pass","cassette":"spec/fixtures/vcr_cassettes/chat_anthropic_claude-sonnet-4-6_compacts_a_long_conversation_and_bills_the_summarization_pass.yml","notes":"Response starts with a compaction block and tokens include every iteration."}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"count_tokens calls /v1/messages/count_tokens with messages, system, tools, tool choice and thinking. A schema or effort setting is not counted, because output_config is left out of the counted payload.","sources":["s497","s498"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"count_tokens_url, render_count_tokens_payload, COUNT_TOKENS_KEYS"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with anthropic/claude-haiku-4-5 counts the next request as configured","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_anthropic_claude-haiku-4-5_counts_the_next_request_as_configured.yml","notes":"Instructions and tools raise the count."}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no image generation operation. Files that code execution produces do not count.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no image editing operation. Files that code execution produces do not count.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no video generation operation. Files that code execution produces do not count.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no speech generation operation. RubyLLM registers no speech protocol here.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no transcription operation. RubyLLM registers no transcription protocol here.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no streaming transcription operation. RubyLLM registers no transcription protocol here.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no speaker identification operation. RubyLLM registers no transcription protocol here.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no moderation operation. The content moderation guide uses ordinary Messages prompts, and classifier refusals are not a moderation endpoint.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no OCR operation. Asking Claude to read an image or PDF is chat, not an OCR API.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Anthropic says it does not offer its own embedding model and points to Voyage AI, a separate service. RubyLLM raises an error.","sources":["s34"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/embeddings.rb","method":"embed"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Anthropic says it does not offer its own embedding model and points to Voyage AI, a separate service.","sources":["s34"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/embeddings.rb","method":"embed"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no reranking operation. The embeddings guide points to Voyage AI, a separate service.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to /v1/files with expires_in, and file_id attachments reach chat for images, PDFs and text. Requests send files-api-2025-04-14 and get its response shape, which omits expires_at.","sources":["s39"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/files.rb","method":"render_upload_fields, BETA_HEADER, parse_file_response"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"provider_file_attachable?, apply_files_beta"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads anthropic uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_anthropic_uploads_a_pdf_through_the_files_api.yml","notes":"Upload returns the file ID and byte size. The endpoint takes no model."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download and UploadedFile.find reach the file content and metadata endpoints. Anthropic allows downloading files that skills or code execution created, not files you uploaded. No recording covers a download.","sources":["s39"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find, download"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/files.rb","method":"file_headers"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch creates, finds and cancels Message Batches, and returns ordered typed results at half price. Requests in a batch run independently and can name different models, and every beta a request needs joins the batch request.","sources":["s499","s186"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/batches.rb","method":"create_batch, find_batch, cancel_batch, batch_results, batch_betas"},{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"batch_cost_multiplier"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with anthropic/claude-haiku-4-5 reloads a batch by id and collects messages without the chats","cassette":"spec/fixtures/vcr_cassettes/batch_with_anthropic_claude-haiku-4-5_reloads_a_batch_by_id_and_collects_messages_without_the_chats.yml","notes":"Batch created, reloaded by ID and its results collected."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"Anthropic offers no embedding endpoint, and Message Batches carry Messages requests only.","sources":["s34","s499"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/embeddings.rb","method":"embed"}]},"advisor_tool":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"A raw advisor_20260301 definition in with_provider_tools and the advisor-tool-2026-03-01 beta in with_headers reach the API. The server_tool_use and advisor_tool_result blocks replay as returned. RubyLLM has no alias for it.","sources":["s500"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"apply_provider_tools"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"server_tool_block?"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row: Perplexity Agent API.","sources":["s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"agent_skills":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Pass skills in the Messages container parameter with with_provider_options, together with the code execution alias. RubyLLM has no call for the Skills API that uploads and versions custom skills. with_skills is a separate feature that loads local folders.","sources":["s501"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_skills"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Managed Agents sessions are hosted agent runs, not an asynchronous research operation.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Message Batches and background compaction are separate operations.","sources":["s186","s33","s38"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"browser_use":{"v1":"passthrough","v2":"partial","offered":true,"notes":"A raw browser_toolset_20260801 definition works with Ruby tools in the browser namespace: member identity, order and text or image results are kept. browser_state result blocks, which report tabs after navigation and tab management, cannot be returned.","sources":["s502"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS, render_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"skipped_tool_call_error"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"computer_use":{"v1":"passthrough","v2":"native","offered":true,"notes":"Declare computer_toolset_20260801 with with_provider_tools and implement each member as a Ruby tool in the computer namespace. RubyLLM keeps member identity, runs calls in order, answers skipped calls, and returns text and image results. Your application owns the desktop. Opus 4.7, Opus 4.6, Sonnet 4.6 and Opus 4.5 take only the older computer_20251124 tool, which needs a raw definition and its beta header and does not dispatch to namespaced tools.","sources":["s503"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS, render_tools, parse_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"native_tool_namespace?, render_tool_call_with_thinking"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"skipped_tool_call_error"}],"verification":[{"status":"passed","recorded_on":"2026-10-25","model":"claude-sonnet-5","spec":"spec/ruby_llm/chat_tool_namespaces_live_spec.rb","example":"RubyLLM::Chat Anthropic computer toolset returns a screenshot to a computer member call","cassette":"spec/fixtures/vcr_cassettes/chat_anthropic_computer_toolset_returns_a_screenshot_to_a_computer_member_call.yml","notes":"A screenshot member returns an image in a tool result with toolset_name."}]},"context_editing":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Send context_management edits such as clear_tool_uses_20250919 with with_provider_options and the context-management-2025-06-27 beta with with_headers. These edits replace the list with_compaction renders, so put compact_20260112 in the same raw list to use both.","sources":["s504"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_headers"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Anthropic offers no embedding endpoint.","sources":["s34","s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Send speed: fast with with_provider_options. Opus 5.5 needs no beta header. Opus 5 and Opus 4.8 are in research preview and need fast-mode-2026-02-01 through with_headers. Fast mode on Opus 4.8 ends on 2026-11-30. RubyLLM does not apply fast mode prices to Cost.","sources":["s208","s484","s491"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_headers"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"The Files API lists and deletes files. RubyLLM implements upload, metadata lookup and download only, and provider options cannot reach these endpoints.","sources":["s39"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload, find, download"},{"version":"v2","path":"lib/ruby_llm/uploaded_file.rb","method":"upload, find, download"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. The Messages API has no suffix parameter.","sources":["s186","s484"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no fine-tuning operation.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Classifier refusals have no request settings, and Enterprise inference hooks are organization administration.","sources":["s186","s33","s505"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Message threads keep a conversation on the server for a short time, in beta, and Managed Agents sessions keep agent state. Chat history stays in the application.","sources":["s506","s507","s186"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s499"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row: Google Interactions API.","sources":["s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"JSON schema structured output exists. The structured outputs page documents no separate JSON object mode.","sources":["s187"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row: Amazon Bedrock Mantle.","sources":["s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"passthrough","offered":true,"notes":"On-demand compaction is in beta. Send compaction: { type: \"summarize\" } with with_provider_options and compact-2026-09-04 with with_headers, then replace the summarized messages with messages= and keep the header. The returned block replays as raw content. Chat#compact raises on this provider.","sources":["s508","s38"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_headers"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"server_tool_block?, render_raw_assistant_message"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"messages="},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"compaction_url"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Message Batches carry Messages requests only.","sources":["s499","s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row: Google Gemini Omni.","sources":["s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. The direct API serves Claude models only.","sources":["s186"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"prefix_completion":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Older Claude models accept a final assistant message as a prefill. Add one with add_message and call generate. RubyLLM has no dedicated prefill interface. Claude 4.6 and later return a 400 error.","sources":["s509","s491"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"add_message, generate"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"render_messages"}]},"programmatic_tool_calling":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Add allowed_callers to a tool with its provider_options and enable the code execution alias. Tool calls and server blocks replay as returned. The API requires the returned container ID on the next request, which the chat loop does not send, so a before_request hook has to add it. No recording covers this flow.","sources":["s510"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"function_for"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"before_request"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching(ttl:) and cache_until_here(ttl:) set 5 minute or 1 hour lifetimes, and one-hour writes are priced at twice the input rate. Cache diagnostics need the raw diagnostics option.","sources":["s37"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_options, prompt_cache_control"},{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"cache_write_input_multiplier"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat prompt cache round-trip anthropic/claude-haiku-4-5 writes a one-hour boundary and prices it as one","cassette":"spec/fixtures/vcr_cassettes/chat_prompt_cache_round-trip_anthropic_claude-haiku-4-5_writes_a_one-hour_boundary_and_prices_it_as_one.yml","notes":"One-hour cache write billed at its own rate."}]},"provider_server_fallback":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Server-side fallback is in beta on the Claude API. Send fallbacks: \"default\" with with_provider_options and server-side-fallback-2026-07-01 with with_headers. with_fallbacks is RubyLLM's own client retry, a separate feature.","sources":["s505"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_headers"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_fallbacks"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The Claude API inventory has no realtime session operation.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Prompt caching reuses a prefix and still generates a reply.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. The text editor tool is a client tool, recorded under shell and computer tools.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"server_tool_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_tools(defer: true), Tool.defer and with_mcp(defer: true) mark tools defer_loading and add tool_search_tool_bm25_20251119. A raw regex search tool replaces the default. Search blocks replay, and history drops them when a request declares no search tool.","sources":["s40"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"NATIVE_TOOL_SEARCH, apply_tool_search, deferred?"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_tools, with_mcp"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"tool_search?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"claude-haiku-4-5-20251001","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with anthropic/claude-haiku-4-5 loads a deferred tool through tool search and calls it","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_anthropic_claude-haiku-4-5_loads_a_deferred_tool_through_tool_search_and_calls_it.yml","notes":"Model searches, loads and calls a deferred tool."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"passthrough","v2":"partial","offered":true,"notes":"The computer toolset works through Ruby tools in the computer namespace. The bash and text editor tools need a raw definition and have no member dispatch: a Ruby tool of the same name renders a second declaration, so a before_request hook has to drop one. Your application runs every action.","sources":["s511","s512","s503"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"render_tools, function_for"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_tools, before_request"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"task_budgets":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Send output_config.task_budget with with_provider_options and task-budgets-2026-03-13 with with_headers. It merges with the effort that with_thinking sets. The budget is advisory, and the chat loop does not track it.","sources":["s217"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/anthropic.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options, with_headers"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"add_thinking_fields"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Managed Agents memory stores hold text documents and are not vector stores.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory. Web fetch is a Messages server tool, not a standalone endpoint.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited API inventory.","sources":["s186","s33"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/anthropic.rb","method":"protocol"}]}},"gaps":[{"notes":"File listing and deletion: The Files API lists and deletes files. RubyLLM has no call for either, and provider options cannot reach them.","sources":["s39"]},{"notes":"Browser use: RubyLLM keeps browser member calls in order and returns text and images, but cannot return the browser_state blocks that report tabs after navigation and tab management.","sources":["s502"]},{"notes":"Shell and computer tools: The computer toolset is built in. The bash and text editor tools need a raw definition, and a Ruby tool of the same name adds a second declaration unless a before_request hook drops one.","sources":["s511","s512","s503"]},{"notes":"Computer use: Opus 4.7, Opus 4.6, Sonnet 4.6 and Opus 4.5 take only the older computer_20251124 tool, which needs a raw definition and does not dispatch to namespaced tools.","sources":["s503"]},{"notes":"Manual context compaction: On-demand compaction needs raw options and a manual swap of the summarized messages. Chat#compact raises on this provider.","sources":["s508"]},{"notes":"Advisor tool, Agent skills, Context editing, Fast inference, Programmatic tool calling, Provider-managed fallback, Task budgets: These need provider options, raw tool definitions or beta headers. RubyLLM has no named interface. The Skills API that uploads custom skills has no RubyLLM call, and Cost does not apply fast mode prices.","sources":["s33","s208","s501"]},{"notes":"Prefix completion: Older Claude models accept a final assistant message. RubyLLM has no dedicated interface, and Claude 4.6 and later return a 400 error.","sources":["s509"]},{"notes":"Tool choice: Opus 5.5, Sonnet 5.5 and Fable 5.1 reject the any and tool choices with a 400 error.","sources":["s490","s491"]},{"notes":"Thinking controls: Per-message effort changes and the progress updates display beta have no named interface.","sources":["s489","s488"]},{"notes":"Token counting: The counted payload leaves out output_config, so a schema or effort setting is not counted, though the endpoint accepts it.","sources":["s498"]}]},{"id":"gemini","name":"Gemini","notes":"Scope: the Gemini Developer API, with generateContent as the default route and Interactions selected by protocol: :interactions, plus media, files, caches, batches and one-shot Live transcription. Live conversations, stored conversations and other Google Cloud products are outside this inventory.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"generateContent is the default route; protocol: :interactions selects Interactions. Google calls generateContent a legacy API that stays supported for existing integrations and recommends Interactions for new work. Parameters vary by model: Gemini 3 and later drop sampling parameters.","sources":["s513","s514","s515","s516","s517"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"render_payload, parse_completion_body"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload, parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality gemini/gemini-2.5-flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_gemini_gemini-2_5-flash_can_have_a_basic_conversation.yml","notes":"Asserts a real answer from generateContent."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Streams through streamGenerateContent on the default route and through server-sent Interactions events on protocol: :interactions.","sources":["s518","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/streaming.rb","method":"build_chunk"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/streaming.rb","method":"stream_response, build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses gemini/gemini-2.5-flash supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_gemini_gemini-2_5-flash_supports_streaming_responses.yml","notes":"Request uses streamGenerateContent with alt=sse."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends responseJsonSchema on generateContent and response_format on Interactions. Batches send the same schema as responseSchema.","sources":["s519","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"structured_output_config, build_json_schema"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/batches.rb","method":"batch_schema_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3-flash-preview","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat#with_schema with gemini/gemini-3-flash-preview accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_gemini_gemini-3-flash-preview_accepts_a_json_schema_and_returns_structured_output.yml","notes":"Asserts JSON matching the schema."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Images go as chat attachments, inline or by Files API reference, on both routes.","sources":["s520"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models gemini/gemini-2.5-flash can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_gemini_gemini-2_5-flash_can_understand_local_images.yml","notes":"Asserts the reply describes the image."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDFs go as chat attachments on both routes. Google documents up to 1000 pages.","sources":["s521"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models gemini/gemini-2.5-flash understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_gemini_gemini-2_5-flash_understands_pdfs.yml","notes":"Asserts the reply reads the PDF."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Audio goes as a chat attachment on both routes.","sources":["s522"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models gemini/gemini-2.5-flash can understand audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_gemini_gemini-2_5-flash_can_understand_audio.yml","notes":"Asserts the reply describes the audio."}]},"video_input":{"v1":"native","v2":"native","offered":true,"notes":"Video goes as a chat attachment on both routes, with a processing mode for agentic navigation on the models that support it.","sources":["s523"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models gemini/gemini-2.5-flash can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_gemini_gemini-2_5-flash_can_understand_local_videos.yml","notes":"Asserts the reply describes the video."},{"status":"passed","recorded_on":"2026-10-14","model":"gemini-3.5-flash-lite","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat agentic video gemini/gemini-3.5-flash-lite navigates a video and answers a follow-up","cassette":"spec/fixtures/vcr_cassettes/chat_agentic_video_gemini_gemini-3_5-flash-lite_navigates_a_video_and_answers_a_follow-up.yml","notes":"Covers navigation steps and the replayed follow-up."}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Returns thought summaries and signatures from both routes and replays signatures in later requests.","sources":["s524","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_thought_parts, parse_thought_signature"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"parse_interaction_thinking"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3-flash-preview","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking gemini/gemini-3-flash-preview returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_gemini_gemini-3-flash-preview_returns_thinking_when_available.yml","notes":"Asserts thinking text in the response."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking maps effort to thinkingLevel and a budget to thinkingBudget on generateContent. Google replaces budgets with levels on Gemini 3.6 and later; a bare with_thinking sends a level on models whose registry entry lists efforts, and -1 only on toggle models such as gemini-2.5-flash. An explicit budget: is still sent as thinkingBudget. Interactions takes effort and summary display only, and RubyLLM raises for a budget or thinking off there.","sources":["s524","s517","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"build_thinking_config"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_interaction_thinking"}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"Maps grounding metadata and Interactions annotations to Citation values. Citations come from a grounding tool such as with_provider_tools(:web_search); with_citations alone does not add a tool.","sources":["s525","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_citations, support_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"parse_interaction_citations"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with gemini/gemini-2.5-flash returns grounding citations when search is enabled","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_gemini_gemini-2_5-flash_returns_grounding_citations_when_search_is_enabled.yml","notes":"Asserts Citation values from grounding metadata."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools work on both routes, including per-tool thought signatures and media in tool results on Gemini 3 and later. Tool results can carry inline media or a Files API reference next to the response.","sources":["s526","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"render_tools, parse_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/tools.rb","method":"render_interaction_tools, parse_interaction_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling gemini/gemini-2.5-flash can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_gemini_gemini-2_5-flash_can_use_tools.yml","notes":"Round trip of a function call and its result."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps auto, none, required and a specific tool to functionCallingConfig on generateContent and tool_choice on Interactions.","sources":["s526","s515","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"build_tool_config"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/tools.rb","method":"render_interaction_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control gemini/gemini-2.5-flash respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_gemini_gemini-2_5-flash_respects_specific_tool_choice.yml","notes":"Request carries allowedFunctionNames."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Receives several function calls in one response and returns results in call order, because functionResponse parts pair by position.","sources":["s526"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/tools.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"parse_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"in_call_order"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling gemini/gemini-2.5-flash can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_gemini_gemini-2_5-flash_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Response carries more than one functionCall."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"Neither the function calling guide nor the generateContent and Interactions references document a switch that limits parallel calls.","sources":["s526","s515","s516"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"build_tool_config"}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"web_search and google_search aliases on both routes, with queries and pages on the server tool call and billed searches counted.","sources":["s525","s527"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES, parse_web_sources"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with gemini/gemini-3.5-flash grounds the answer and exposes the queries it ran","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_gemini_gemini-3_5-flash_grounds_the_answer_and_exposes_the_queries_it_ran.yml","notes":"Response carries webSearchQueries and grounding chunks."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions streams a Google Search answer with suggestions it does not keep for replay","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_streams_a_google_search_answer_with_suggestions_it_does_not_keep_for_replay.yml","notes":"Covers the Interactions route."}]},"server_web_fetch":{"v1":"passthrough","v2":"native","offered":true,"notes":"web_fetch and url_context aliases on both routes. Google limits which URL types and models work.","sources":["s528","s527"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"metadata_server_tool_calls"}]},"server_code_execution":{"v1":"passthrough","v2":"native","offered":true,"notes":"code_execution alias on both routes. Returns execution parts as server tool calls and replays them in later turns.","sources":["s529","s527"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_server_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat code execution with gemini/gemini-3.5-flash runs code server-side and replays the turn","cassette":"spec/fixtures/vcr_cassettes/chat_code_execution_with_gemini_gemini-3_5-flash_runs_code_server-side_and_replays_the_turn.yml","notes":"Response carries executableCode and codeExecutionResult parts."}]},"server_file_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"file_search alias on both routes uses an existing store and parses grounding. Creating and filling stores is the separate store management row; RubyLLM.upload targets ordinary Files.","sources":["s530","s527"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb","method":"SERVER_TOOL_ALIASES"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(mcp: {name:, url:}) on protocol: :interactions connects a Streamable HTTP server. Google documents remote MCP for Interactions. Calls, results and signed steps replay statelessly. Remote tools run without an approval step. The generateContent schema also lists mcpServers; RubyLLM has no alias for it there.","sources":["s526","s516","s515"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/tools.rb","method":"parse_interaction_server_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_interaction_history"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions executes a remote MCP tool and replays its signed results through stateless chat","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_executes_a_remote_mcp_tool_and_replays_its_signed_results_through_stateless_chat.yml","notes":"Fresh Microsoft Learn MCP execution returned actual call/result steps, then a second user turn used those results. MCP output signatures remain in raw history but are omitted from input as required by the input schema; thought signatures are retained."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions streams remote MCP results and preserves the complete signed history","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_streams_remote_mcp_results_and_preserves_the_complete_signed_history.yml","notes":"Fresh SSE execution reconstructed call, result, thought and model-output steps; text chunks equal final assistant text and actual prompt/tool/thought usage is normalized."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Implicit caching applies automatically on Gemini 2.5 and later on both routes, and RubyLLM reports cached tokens. An ineligible prompt stays uncached; thresholds vary by model.","sources":["s531"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb","method":"parse_completion_response"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"maybe_log_implicit_caching_note, parse_completion_body"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"parse_interaction_usage"}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Gemini has no cache breakpoint in a request. RubyLLM accepts cache_until_here and only logs a note, because caching is implicit or goes through cache resources.","sources":["s531","s515","s532"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"maybe_log_implicit_caching_note"}]},"explicit_cache":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.cache creates, finds, updates and deletes cachedContents, and with_caching(id:) attaches one on generateContent. Google does not offer explicit caching on Interactions or on Gemini 4 Argon and later; RubyLLM raises for a cache id on Interactions.","sources":["s531","s532","s533"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/caches.rb"},{"version":"v2","path":"lib/ruby_llm/cached_content.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"ensure_interaction_carries"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/cached_content_spec.rb","example":"RubyLLM::CachedContent explicit caching round-trip gemini/gemini-2.5-flash creates, uses, extends, and deletes a cache","cassette":"spec/fixtures/vcr_cassettes/cachedcontent_explicit_caching_round-trip_gemini_gemini-2_5-flash_creates_uses_extends_and_deletes_a_cache.yml","notes":"Full lifecycle against cachedContents."}]},"compaction":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Gemini documents context window compression only for Live sessions. This is outside RubyLLM 2.0 coverage; conversation state stays in the application. with_compaction is dropped on generateContent.","sources":["s534","s535"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"chat.count_tokens calls countTokens with rendered contents, instructions and tools on generateContent. Interactions has no RubyLLM token counting method.","sources":["s536","s537"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"render_count_tokens_payload, count_tokens_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-flash","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with gemini/gemini-3.5-flash counts the next request as configured","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_gemini_gemini-3_5-flash_counts_the_next_request_as_configured.yml","notes":"Asserts a positive count from countTokens."}]},"image_generation":{"v1":"native","v2":"native","offered":true,"notes":"paint uses generateContent for Gemini image models, with multiple results, aspect ratio and resolution. Google shut down Imagen, so the predict path has no model. Gemini Nano Banana 2.1 does not match the image model test and goes to predict.","sources":["s538","s539","s540","s541"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/images.rb","method":"render_gemini_image_payload, gemini_image_model?, image_endpoint_action"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.1-flash-lite-image","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality gemini/gemini-3.1-flash-lite-image can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_gemini_gemini-3_1-flash-lite-image_can_paint_images.yml","notes":"Request goes to generateContent and returns an image."}]},"image_editing":{"v1":"partial","v2":"native","offered":true,"notes":"paint accepts reference images through with: on Gemini image models. The guide edits by prompt and has no mask field; RubyLLM raises for mask:. Gemini Nano Banana 2.1 has the same routing limit as image generation.","sources":["s538"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/images.rb","method":"render_image_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/images.rb","method":"render_gemini_image_payload, validate_paint_inputs!, image_attachments"}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate runs Gemini Omni through a stored background interaction and returns a Video with its billed tokens. VideoJob.find picks the job back up. Google shut down Veo in the Gemini API on 2026-10-22.","sources":["s542","s543","s544","s541"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/videos.rb","method":"video_url, render_video_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"parse_interaction_part"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak uses generateContent with a voice name and returns Speech. Multi-speaker and output format settings go through provider_options. Gemini 3.8 TTS returns WAV by default, yet RubyLLM labels the result as pcm. A stored voice ID belongs in voiceConfig.voice, while voice: writes prebuiltVoiceConfig. Earlier TTS models may shut down from 2026-11-17.","sources":["s545","s546","s515","s541"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/speech.rb","method":"render_speech_payload, parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash-preview-tts","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality gemini/gemini-2.5-flash-preview-tts can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_gemini_gemini-2_5-flash-preview-tts_can_speak.yml","notes":"Request goes to generateContent and returns audio."}]},"transcription":{"v1":"native","v2":"native","offered":true,"notes":"transcribe uses generateContent audio understanding on general models and returns a typed transcript. The dedicated model is its own row.","sources":["s522","s547"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/transcription.rb","method":"transcribe, render_transcription_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality gemini/gemini-2.5-flash can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_gemini_gemini-2_5-flash_can_transcribe_audio.yml","notes":"Asserts text from generateContent."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"transcribe with gemini-3.5-transcribe-live streams BidiGenerateContent partial and final text into TranscriptionChunk and returns a Transcription. It takes mono 16-bit PCM WAV with manual activity boundaries, and rejects diarization and word timestamps, which Google does not offer on Live. One-shot transcription is kept; live conversations are outside scope.","sources":["s547","s548"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/live_transcription.rb","method":"transcribe, render_transcription_setup"}],"verification":[{"status":"passed","recorded_on":"2026-09-17","model":"gemini-3.5-transcribe-live","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription streams partial and final transcription through gemini Live","cassette":"spec/fixtures/websocket_cassettes/transcription_google_gemini.json","notes":"Fresh provider execution and offline replay passed. Actual WebSocket setup, partial/final transcription and generation-complete frames."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"The dedicated gemini-3.5-transcribe model takes speaker_names: [] to request detected speaker labels, and timestamps: :word keeps numeric timing. Google documents up to 8 speakers, with 3 or more experimental. Named speakers and reference clips are rejected. Live transcription has no diarization.","sources":["s547","s548"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/interactions/transcription.rb","method":"render_transcription_payload, parse_transcription_word"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/transcription.rb","method":"rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription transcribes audio with speaker labels and word timestamps through gemini","cassette":"spec/fixtures/vcr_cassettes/transcription_transcribes_audio_with_speaker_labels_and_word_timestamps_through_gemini.yml","notes":"Fresh provider execution and offline replay passed. Typed transcript, numeric word offsets and actual detected speaker labels; the two-speaker fixture returned distinct labels."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription distinguishes two speakers through gemini dedicated transcription","cassette":"spec/fixtures/vcr_cassettes/transcription_distinguishes_two_speakers_through_gemini_dedicated_transcription.yml","notes":"Fresh provider execution and offline replay passed. Typed transcript, numeric word offsets and actual detected speaker labels; the two-speaker fixture returned distinct labels."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no moderation endpoint. Safety settings on a generation request are the guardrail row.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no OCR or document parsing endpoint. Reading a PDF with a chat model is PDF input.","sources":["s549","s533","s521"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"embed uses batchEmbedContents with dimensions, task_type and title.","sources":["s550","s551"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/embeddings.rb","method":"embedding_url, render_embedding_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-embedding-001","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality gemini/gemini-embedding-001 can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_gemini_gemini-embedding-001_can_handle_a_single_text.yml","notes":"Asserts a vector from batchEmbedContents."}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"embed renders image, audio, video and PDF parts for gemini-embedding-2. RubyLLM raises when an array of texts is combined with attachments.","sources":["s550","s551"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/embeddings.rb","method":"embedding_media?, media_embedding_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-embedding-2","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings gemini/gemini-embedding-2 embeds an image alongside text","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_gemini_gemini-embedding-2_embeds_an_image_alongside_text.yml","notes":"Request carries inline image data."}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no reranking endpoint. Asking a chat model to rank text is chat.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload uses the resumable Files API and waits for processing, then chat attachments reference the file. Registering Cloud Storage files through files:register is not implemented.","sources":["s552","s553"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/files.rb","method":"upload, start_resumable_upload, await_active"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":null,"spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads gemini uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_gemini_uploads_a_pdf_through_the_files_api.yml","notes":"Covers the upload and file metadata."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download follows a file's downloadUri. Google does not let you download user-uploaded files, only files models generate, and RubyLLM raises when a file has no URI.","sources":["s552","s553"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/files.rb","method":"download"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"Batch.submit sends inline generateContent requests for one model, with polling, cancellation and ordered results. Google recommends the JSONL input file for larger jobs, since inline requests stop at 20 MB; RubyLLM does not submit input files.","sources":["s554","s555"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/batches.rb","method":"create_batch, batch_results, cancel_batch"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with gemini/gemini-2.5-flash answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_gemini_gemini-2_5-flash_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Submission through batchGenerateContent and delivered answers."}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"embed_later and Batch submit inline text requests to asyncBatchEmbedContent. Lookup and cancellation share the Batch lifecycle, and item metadata restores scalar and array result shapes in order. Input files and staged multimodal inputs are not implemented.","sources":["s554","s555","s551"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/embedding_batches.rb","method":"embedding_batch_requests, parse_embedding_batch_results"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/batches.rb","method":"create_batch, batch_results"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-embedding-001","spec":"spec/ruby_llm/protocols/gemini/embedding_batches_spec.rb","example":"RubyLLM::Protocols::Gemini::EmbeddingBatches submits, retrieves, and cancels an asynchronous embedding batch","cassette":"spec/fixtures/vcr_cassettes/protocols_gemini_embeddingbatches_submits_retrieves_and_cancels_an_asynchronous_embedding_batch.yml","notes":"Live submit, find and cancel."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory documents no advisor tool.","sources":["s549","s533","s527"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row is the Perplexity Agent API, which is a different service.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"agent_skills":{"v1":"missing","v2":"missing","offered":true,"notes":"Gemini managed agents mount SKILL.md files into a remote environment. RubyLLM has no agent or environment resources on Gemini, and chat requests name a model, not an agent.","sources":["s556","s557","s533"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"}]},"async_research":{"v1":"missing","v2":"missing","offered":true,"notes":"Deep Research is a preview agent that runs in the background through Interactions. RubyLLM.research raises on this provider because it declares no research protocol. Vertex AI has its own.","sources":["s558","s559"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb","method":"protocol :gemini"},{"version":"v2","path":"lib/ruby_llm/research_job.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row concerns the Responses protocol, which Gemini does not serve. Interactions has its own background execution, covered by the research row.","sources":["s549","s533","s560","s559"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"browser_use":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"The Computer Use tool targets browser environments and is generally available. A raw tool Hash reaches both routes through with_provider_tools. Your application runs the actions and sends back screenshots.","sources":["s561"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no classification endpoint. Prompting a chat model is chat.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"computer_use":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Computer Use is generally available on both APIs. RubyLLM has no alias for it. A raw tool Hash, such as computerUse on generateContent or type computer_use on Interactions, is forwarded, and your application runs the returned actions.","sources":["s561","s515","s562"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"}]},"context_editing":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Live contextWindowCompression can drop old turns through a sliding window. Gemini documents this only for Live sessions, which are outside RubyLLM 2.0 coverage.","sources":["s534","s535"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no contextualized embedding operation. task_type and title are ordinary embedding fields.","sources":["s549","s533","s550"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"transcribe with gemini-3.5-transcribe goes through Interactions with transcription_config and returns text, usage, word offsets and detected speakers. Google rejects custom vocabulary combined with diarization or word timestamps.","sources":["s547","s516"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/transcription.rb","method":"transcription_url, render_transcription_payload, parse_transcription_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription transcribes audio with speaker labels and word timestamps through gemini","cassette":"spec/fixtures/vcr_cassettes/transcription_transcribes_audio_with_speaker_labels_and_word_timestamps_through_gemini.yml","notes":"Fresh provider execution and offline replay passed. Typed transcript, numeric word offsets and actual detected speaker labels; the two-speaker fixture returned distinct labels."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription distinguishes two speakers through gemini dedicated transcription","cassette":"spec/fixtures/vcr_cassettes/transcription_distinguishes_two_speakers_through_gemini_dedicated_transcription.yml","notes":"Fresh provider execution and offline replay passed. Typed transcript, numeric word offsets and actual detected speaker labels; the two-speaker fixture returned distinct labels."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no dubbing endpoint. Live translate is a realtime session.","sources":["s549","s533","s563"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Priority inference is selected with service_tier in the request body. with_provider_options(service_tier: \"priority\") forwards it. Flex uses the same field.","sources":["s564","s565","s515"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Google documents listing and deleting Files. RubyLLM has upload, find and download only.","sources":["s552","s553"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/files.rb","method":"upload, download"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Google documents File Search store and document lifecycles and import. RubyLLM can only name an existing store on the file_search tool.","sources":["s530","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/files.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no completion endpoint with a suffix. The OpenAI compatibility layer lists chat completions, images, embeddings and batches.","sources":["s549","s533","s560"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"Google states no Gemini API model supports fine-tuning and has no plans to add it. Tuning on Google Cloud is a separate product.","sources":["s566"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"google_maps":{"v1":"passthrough","v2":"native","offered":true,"notes":"google_maps alias on both routes. Options such as a location go in the alias hash. Availability depends on the model.","sources":["s567","s527"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb","method":"SERVER_TOOL_ALIASES"}]},"guardrail_configuration":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"safetySettings pass through with_provider_options on generateContent. RubyLLM has no shared API for thresholds, and it maps blocks to content_filter.","sources":["s568","s515"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"FINISH_REASONS"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Interactions can store requests and continue them with previous_interaction_id. This is outside RubyLLM 2.0 coverage; chat sends store: false and replays application-owned history.","sources":["s569","s516"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"}]},"image_batches":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Google runs image generation through Batch. A staged chat with raw image response modalities is forwarded and its attachments are parsed. There is no batched paint.","sources":["s538","s554"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/batches.rb","method":"create_batch"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_completion_body"}]},"interactions_api":{"v1":"missing","v2":"native","offered":true,"notes":"protocol: :interactions handles text and media input, local tools, JSON Schema output, streaming, citations, thinking and usage. It sends store: false and replays results from application-owned history. Stored continuation is outside scope. Omni, Lyria and agents have no dedicated RubyLLM operation on it. Combining local tools with with_schema on this protocol can loop: on November 10, 2026 the model kept calling the tool after each result.","sources":["s569","s516","s570","s526"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb","method":"protocol :interactions"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/tools.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions executes a remote MCP tool and replays its signed results through stateless chat","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_executes_a_remote_mcp_tool_and_replays_its_signed_results_through_stateless_chat.yml","notes":"Fresh Microsoft Learn MCP execution returned actual call/result steps, then a second user turn used those results. MCP output signatures remain in raw history but are omitted from input as required by the input schema; thought signatures are retained."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions streams remote MCP results and preserves the complete signed history","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_streams_remote_mcp_results_and_preserves_the_complete_signed_history.yml","notes":"Fresh SSE execution reconstructed call, result, thought and model-output steps; text chunks equal final assistant text and actual prompt/tool/thought usage is normalized."},{"status":"unavailable","checked_on":"2026-11-10","model":"gemini-3.8-flash","notes":"With both a local tool and a JSON Schema, Gemini called the tool again after every result and never completed; gemini-3.7-flash and gemini-3.1-pro-preview did the same. Tools alone and a schema alone each completed. The recording from October 11, 2026, which passed, was removed."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.8-flash","spec":"spec/ruby_llm/protocols/interactions_spec.rb","example":"RubyLLM::Protocols::Interactions streams local function arguments and continues with the actual tool result","cassette":"spec/fixtures/vcr_cassettes/protocols_interactions_streams_local_function_arguments_and_continues_with_the_actual_tool_result.yml","notes":"Fresh function-call argument deltas start after an empty object, assemble into an object for replay, execute multiply and return323."}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"generationConfig.responseMimeType of application/json without a schema passes through with_provider_options.","sources":["s515","s519"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row is Amazon Bedrock Mantle, which is a different service.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"Live offers automatic window compression only, with no separate compact operation.","sources":["s534","s535"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"music_generation":{"v1":"partial","v2":"passthrough","offered":true,"notes":"The music guide generates Lyria audio through Interactions. A chat on protocol: :interactions with a Lyria model returns the audio as an attachment. RubyLLM has no music operation, and Lyria RealTime needs a Live session. No recording covers it.","sources":["s571","s572","s573"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini/chat.rb","method":"parse_completion_response"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"parse_interaction_part"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"}]},"nonchat_batches":{"v1":"missing","v2":"passthrough","offered":true,"notes":"TTS models list Batch API support. A staged chat with raw AUDIO response modalities is forwarded and its attachments are parsed. There is no batched speak.","sources":["s546","s574","s554"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/batches.rb","method":"create_batch"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_completion_body"}]},"omni_video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate submits Omni jobs through Interactions, with aspect ratio, duration and resolution through response_format in provider_options. A chat on protocol: :interactions also returns the video as an attachment.","sources":["s544","s575","s542","s541"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"parse_interaction_part"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/chat.rb","method":"render_payload"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory serves Google models only.","sources":["s549","s533","s513"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"prefix_completion":{"v1":"na","v2":"passthrough","offered":true,"notes":"Gemini 2.5 Flash and 3.1 Flash-Lite continue a final model turn; gemini-3.6-flash returned 400, and the migration guides list prefill as removed on 3.6 and later. Add the partial answer with add_message and call generate. RubyLLM has no dedicated prefill interface.","sources":["s517","s576","s514","s515"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"render_messages"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"add_message, generate"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"gemini-3.1-flash-lite","notes":"Through add_message and generate, the model continued a prefilled \"Zaffre, then\" in place. gemini-2.5-flash continued it through raw requests. gemini-3.6-flash returned 400: Requests ending with a model turn are not supported."}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory and tool guides do not document code that calls tools programmatically.","sources":["s549","s533","s527","s577","s529"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching(id:) attaches a cache resource on generateContent, and RubyLLM.cache sets its TTL. Implicit caching takes no options. This does not apply to Interactions or Gemini 4 Argon.","sources":["s531","s532"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/caches.rb","method":"render_cache_payload, render_cache_ttl"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory documents no provider-managed model fallback. Google reroutes retired model names, which is a deprecation policy.","sources":["s549","s533","s541"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"The Live API documents realtime voice and video sessions. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s563","s535","s578"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory documents no cache of whole responses.","sources":["s549","s533","s531"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"Gemini serves no Responses endpoint. The OpenAI compatibility layer lists chat completions, images, embeddings and batches; Interactions and Live are separate APIs.","sources":["s549","s533","s560"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"Gemini serves no Responses endpoint, so it has no Responses batches.","sources":["s549","s533","s560"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no standalone search endpoint. Grounding is a tool on generation requests.","sources":["s549","s533","s525"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The audited tools list has no apply patch tool.","sources":["s549","s533","s527","s570"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited tools list has no image generation tool. Image generation is a model operation.","sources":["s549","s533","s527","s570"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The audited tools list has no hosted tool search. RubyLLM gives Gemini chats its own client-side tool search.","sources":["s549","s533","s527","s570"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The audited tools list has no X search tool.","sources":["s549","s533","s527","s570"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"shell_computer_tools":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Computer Use supports browser, mobile and desktop actions through a raw tool definition, and your application executes them. Antigravity agents have a hosted shell, which RubyLLM does not reach.","sources":["s561","s570","s557"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no sound effects endpoint. Lyria generates music.","sources":["s549","s533","s571"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"streaming_speech":{"v1":"missing","v2":"missing","offered":true,"notes":"Gemini streams TTS audio as raw PCM through Interactions. RubyLLM.speak makes one generateContent request, and a block raises because this protocol has no speech streaming.","sources":["s545"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/speech.rb","method":"speech_url"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory documents no task budget.","sources":["s549","s533","s516"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no text analysis endpoint.","sources":["s549","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no general vector store API. File Search stores have their own row.","sources":["s549","s533","s530"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Batch API documents generateContent and embedding jobs, and the Omni guide documents no batch route.","sources":["s554","s575","s533"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"video_editing":{"v1":"missing","v2":"native","offered":true,"notes":"animate takes a video through with: and Omni edits it. Google refuses uploaded video in the EEA, Switzerland and the UK, and the request fails there. Stored conversational edits are outside scope.","sources":["s544"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment, parse_interaction_part"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/videos.rb"}]},"video_extension":{"v1":"missing","v2":"native","offered":true,"notes":"animate(extend:) continues a generated Video through previous_interaction_id, in every region, adding 3 to 10 seconds up to 30 in total. A file or URL also works, except in the EEA, Switzerland and the UK.","sources":["s544","s543","s541"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gemini.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/videos.rb","method":"render_video_extension_payload, render_video_extension"},{"version":"v2","path":"lib/ruby_llm/protocols/interactions/content.rb","method":"render_interaction_attachment, parse_interaction_part"}]},"voice_cloning":{"v1":"na","v2":"missing","offered":true,"notes":"Voice replication creates a voice from a source recording and a consent recording through POST /v1beta/voices, or a stateless voice key. RubyLLM has no voice lifecycle, and speak can only use a voice prepared elsewhere through provider_options.","sources":["s579","s580","s545"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gemini/speech.rb","method":"render_speech_payload"},{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no speech-to-speech conversion. Voice remixing edits a stored voice from a text instruction.","sources":["s549","s533","s581","s580"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory has no standalone fetch endpoint. URL context is a tool on generation requests.","sources":["s549","s533","s528"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"This row is Responses over WebSocket. Gemini serves no Responses endpoint, and its WebSockets belong to Live.","sources":["s549","s533","s535"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gemini.rb"}]}},"gaps":[{"notes":"Image generation, Image editing: paint sends gemini-nano-banana-2.1 to the shut down Imagen predict endpoint, because the model test looks for -image in the id. Other Gemini image models work. Masks are not supported.","sources":["s538","s539","s540"]},{"notes":"Speech generation: speak labels the WAV that Gemini 3.8 TTS returns by default as pcm, rejects other format values, and writes voice: to prebuiltVoiceConfig, so stored voice IDs need provider_options.","sources":["s545","s546","s515"]},{"notes":"Agent skills: Gemini managed agents mount skills into remote environments. RubyLLM has no agent or environment resources on this provider.","sources":["s556","s557"]},{"notes":"Asynchronous research: Deep Research runs as a background Interactions agent. This provider declares no research protocol, so RubyLLM.research raises here.","sources":["s558"]},{"notes":"Browser use, Computer use, Shell and computer tools: Computer Use takes a raw tool definition, and your application executes the actions. RubyLLM has no alias for it.","sources":["s561"]},{"notes":"File listing and deletion: Google documents listing and deleting Files. RubyLLM has upload, find and download only.","sources":["s552"]},{"notes":"File search store management: RubyLLM names an existing store on the file_search tool, but does not create or delete stores or import documents.","sources":["s530"]},{"notes":"Chat batches, Embedding batches: Batches submit inline requests only, which Google limits to 20 MB. JSONL input files are not implemented.","sources":["s554"]},{"notes":"Fast inference, Guardrail configuration, JSON object mode: service_tier, safetySettings and responseMimeType without a schema need raw provider options.","sources":["s564","s568","s515"]},{"notes":"Image batches, Other batch operations: Image and TTS batches work as staged chats with raw response modalities. There is no batched paint or speak.","sources":["s554","s546"]},{"notes":"Music generation: Lyria runs through a chat on Interactions with audio attachments. RubyLLM has no music operation or Lyria RealTime session.","sources":["s571"]},{"notes":"Streaming speech generation: Gemini streams TTS through Interactions. speak has no streaming implementation here.","sources":["s545"]},{"notes":"Voice cloning: Voice replication needs POST /v1beta/voices with consent audio. RubyLLM has no voice lifecycle.","sources":["s579","s580"]},{"notes":"Prefix completion: Gemini 2.5 Flash and 3.1 Flash-Lite continue a final model turn, gemini-3.6-flash returned 400, and RubyLLM has no dedicated prefill interface.","sources":["s517"]}]},{"id":"vertexai","name":"Vertex AI","notes":"The audit covers Gemini Enterprise Agent Platform inference at aiplatform.googleapis.com, including Gemini, Google-managed Claude, Mistral, MaaS partner models, media, caching, batches and hosted research. Anthropic-managed Claude on claude.googleapis.com, Cloud Vision, Document AI and Model Armor are separate services; Discovery Engine ranking and Cloud Text-to-Speech are included because RubyLLM routes to them.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.chat reaches Vertex AI through protocol routing. Gemini uses generateContent, Claude uses rawPredict, Mistral uses its rawPredict dialect, and publisher-prefixed MaaS models use the OpenAI-compatible endpoint. Model restrictions apply.","sources":["s582","s583","s584","s585","s586"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/mistral.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/chat_completions.rb","method":"completion_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality vertexai/gemini-2.5-flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_vertexai_gemini-2_5-flash_can_have_a_basic_conversation.yml","notes":"generateContent answer on Vertex AI."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Block calls stream through Gemini streamGenerateContent, Claude streamRawPredict, Mistral streamRawPredict, and MaaS Chat Completions.","sources":["s583","s584","s585","s586"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"stream_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb","method":"stream_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/mistral.rb","method":"stream_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses vertexai/gemini-2.5-flash supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_vertexai_gemini-2_5-flash_supports_streaming_responses.yml","notes":"Server-sent chunks from streamGenerateContent."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends responseJsonSchema on Gemini, output_config on Claude and response_format on MaaS and Mistral. Model restrictions apply.","sources":["s587","s588","s589"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"structured_output_config"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_output_config"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3-flash-preview","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat with_schema with vertexai/gemini-3-flash-preview accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_vertexai_gemini-3-flash-preview_accepts_a_json_schema_and_returns_structured_output.yml","notes":"responseJsonSchema request and JSON reply."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Chat attachments carry images on Gemini, Claude and vision MaaS models. Local files and uploaded Cloud Storage files work on Gemini. Google-managed Claude does not document URL image sources, while the Anthropic protocol sends URL attachments as URL sources.","sources":["s590","s591"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"url?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models vertexai/gemini-2.5-flash can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_vertexai_gemini-2_5-flash_can_understand_local_images.yml","notes":"Inline image part on generateContent."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"Chat attachments carry PDFs on Gemini and Claude. Google-managed Claude does not document URL document sources or the Files API, so send those PDFs as local files.","sources":["s592","s591"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"url?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models vertexai/gemini-2.5-flash understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_vertexai_gemini-2_5-flash_understands_pdfs.yml","notes":"Inline PDF part on generateContent."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Chat attachments carry audio on Gemini models. No recording covers a chat audio attachment on Vertex AI; the same part shape is recorded for transcription.","sources":["s593"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb","method":"render_content"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality vertexai/gemini-2.5-flash can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_vertexai_gemini-2_5-flash_can_transcribe_audio.yml","notes":"Same inline audio part on generateContent, through RubyLLM.transcribe."}]},"video_input":{"v1":"native","v2":"native","offered":true,"notes":"Chat attachments carry video on Gemini models.","sources":["s594"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb","method":"render_content"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models vertexai/gemini-2.5-flash can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_vertexai_gemini-2_5-flash_can_understand_local_videos.yml","notes":"Inline video part on generateContent."}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Message#thinking returns Gemini thought summaries and signatures, Claude thinking blocks, and MaaS reasoning fields where the model returns them.","sources":["s595","s488","s596"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_thought_parts"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_thinking_blocks"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3-flash-preview","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking vertexai/gemini-3-flash-preview returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_vertexai_gemini-3-flash-preview_returns_thinking_when_available.yml","notes":"Thought summary returned with the answer."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking maps to Gemini thinkingConfig and Claude thinking settings. Model restrictions apply.","sources":["s595","s488"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"build_thinking_config"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb","method":"thinking_block_binding?"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"thinking_off_control"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3-flash-preview","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking vertexai/gemini-3-flash-preview returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_vertexai_gemini-3-flash-preview_returns_thinking_when_available.yml","notes":"Request carries the thinking configuration."}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"Message#citations normalizes Gemini grounding supports and Claude document citations. Google-managed Claude documents citations.","sources":["s597","s365","s35"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_citation"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with vertexai/gemini-2.5-flash returns grounding citations when search is enabled","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_vertexai_gemini-2_5-flash_returns_grounding_citations_when_search_is_enabled.yml","notes":"Grounding metadata becomes Citation values."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"with_tools declares function tools on Gemini, Claude, Mistral and MaaS routes. Results go back through each protocol.","sources":["s598","s584","s599","s600"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"render_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"render_tools"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling vertexai/gemini-2.5-flash can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_vertexai_gemini-2_5-flash_can_use_tools.yml","notes":"Tool call and result round trip."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps to Gemini functionCallingConfig, Claude tool_choice and Chat Completions tool_choice.","sources":["s598","s600","s599"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"build_tool_config"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control vertexai/gemini-2.5-flash respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_vertexai_gemini-2_5-flash_respects_specific_tool_choice.yml","notes":"allowedFunctionNames forces the named tool."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool calls in one response are parsed on Gemini, Claude and Chat Completions routes.","sources":["s598","s492"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling vertexai/gemini-2.5-flash can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_vertexai_gemini-2_5-flash_can_use_parallel_tool_calls.yml","notes":"Multi-tool conversation round trip."}]},"parallel_tool_control":{"v1":"missing","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) maps to disable_parallel_tool_use on Claude and parallel_tool_calls on Chat Completions routes such as Grok. The Gemini function-calling guide documents no such setting, and the Gemini protocol sends none.","sources":["s492","s600","s598"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parallel_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb"}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"The web_search alias sends google_search on Gemini. Google-managed Claude documents only the basic web search tool, and it is off by default under an organization policy. The alias inherited from the Anthropic protocol sends web_search_20260318, which that route does not document.","sources":["s597","s601","s493"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with vertexai/gemini-2.5-flash grounds the answer and lists the pages it used","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_vertexai_gemini-2_5-flash_grounds_the_answer_and_lists_the_pages_it_used.yml","notes":"google_search tool and Message#web_sources."}]},"server_web_fetch":{"v1":"passthrough","v2":"native","offered":true,"notes":"The web_fetch and url_context aliases send url_context on Gemini. Google-managed Claude does not offer web fetch; only Anthropic-managed Claude does, on a separate endpoint.","sources":["s602","s494","s603"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"}]},"server_code_execution":{"v1":"passthrough","v2":"native","offered":true,"notes":"The code_execution alias and returned execution parts work on Gemini. Google-managed Claude does not offer code execution.","sources":["s604","s591","s495"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_server_tool_calls"}]},"server_file_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"The file_search alias maps an existing datastore to retrieval.vertexAiSearch on generateContent. retrievedContext grounding becomes Citation values. Store provisioning and ingestion stay outside RubyLLM.","sources":["s365","s202"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"apply_provider_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"parse_citations"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/interactions_spec.rb","notes":"The alias renders the documented retrieval.vertexAiSearch shape. No recording runs against a live datastore.","example":"uses Vertex AI Search retrieval for its named file-search alias"}]},"server_mcp":{"v1":"missing","v2":"partial","offered":true,"notes":"RubyLLM.research connects a remote MCP server through the hosted Deep Research agent. Ordinary chat has no route: the generateContent tool schema has no MCP tool (400 Unknown name mcpServers), the Interactions API rejects an mcp_server tool on Gemini chat models though the same request with google_search succeeds, and Google-managed Claude does not offer the MCP connector. On the recorded research route, the response omits MCP call and result steps, and its citation has no source URL. Hosted MCP exposes no approval policy; approval options raise before submission.","sources":["s205","s203","s591","s202"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/research.rb","method":"SERVER_TOOL_ALIASES, merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/research_job.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deep-research-preview-04-2026","spec":"spec/ruby_llm/protocols/vertexai/research_spec.rb","example":"completes a short hosted research task through an actual public documentation MCP server","cassette":"spec/fixtures/vcr_cassettes/protocols_vertexai_research_completes_a_short_hosted_research_task_through_an_actual_public_documentation_mcp_server.yml","notes":"Completed hosted task through a public Microsoft Learn MCP server, with a provider-attributed citation."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/research_spec.rb","notes":"Rendering and parsing of the hosted MCP tool."},{"status":"unavailable","checked_on":"2026-11-10","model":"gemini-3.5-flash, gemini-3.8-flash","notes":"Live: generateContent with tools[].mcpServers or mcp_servers returned 400 Unknown name. POST interactions with an mcp_server tool returned 400 invalid_request on both models, with several tool shapes and two servers, while google_search on the same route returned 200."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Gemini caches implicitly and reports cached tokens. with_caching maps to Claude cache_control, including its automatic top-level form.","sources":["s214","s605","s37"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/chat.rb","method":"maybe_log_implicit_caching_note"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_control"}]},"cache_boundaries":{"v1":"missing","v2":"native","offered":true,"notes":"cache_until_here maps to cache_control on Google-managed Claude. Gemini has no per-message breakpoints.","sources":["s605","s37","s214"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"cache_boundary?, inject_cache_control"}]},"explicit_cache":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.cache creates, updates, uses and deletes projects/locations/cachedContents resources on Gemini models.","sources":["s214"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"caches_url, cache_name, cache_model_name"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/caches.rb","method":"render_cache_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/cached_content_spec.rb","example":"RubyLLM::CachedContent explicit caching round-trip vertexai/gemini-2.5-flash creates, uses, and deletes a cache","cassette":"spec/fixtures/vcr_cassettes/cachedcontent_explicit_caching_round-trip_vertexai_gemini-2_5-flash_creates_uses_and_deletes_a_cache.yml","notes":"Create, use and delete against cachedContents."}]},"compaction":{"v1":"missing","v2":"native","offered":true,"notes":"with_compaction sends the Claude context_management compaction edit on Google-managed Claude. Claude documents threshold compaction on that route. Gemini and MaaS routes gain nothing from it. No recording covers this route.","sources":["s496","s603"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction, apply_compaction_headers"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"chat.count_tokens calls Gemini countTokens. Claude on Vertex AI raises an error, although Google documents a count-tokens rawPredict endpoint for it. Some Gemini models, such as the TTS models, do not take Count Tokens.","sources":["s606","s607","s608"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"count_tokens_url, render_count_tokens_payload"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb","method":"count_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with vertexai/gemini-2.5-flash counts a staged message","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_vertexai_gemini-2_5-flash_counts_a_staged_message.yml","notes":"countTokens returns the count before generation."}]},"image_generation":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.paint calls generateContent on Gemini image models. The Imagen predict path remains in code, but the current model list names no Imagen model and the older Imagen page lists discontinued endpoints.","sources":["s609","s610","s611","s612"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gemini/images.rb","method":"images_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"images_url"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/images.rb","method":"render_gemini_image_payload, image_endpoint_action"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.1-flash-lite-image","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality vertexai/gemini-3.1-flash-lite-image can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_vertexai_gemini-3_1-flash-lite-image_can_paint_images.yml","notes":"generateContent returns inline image data."}]},"image_editing":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.paint(with:) sends reference images to Gemini image models, the documented editing route. Masks are rejected because Gemini image models take none. No recording covers editing on Vertex AI.","sources":["s613"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"images_url"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/images.rb","method":"validate_paint_inputs!, render_gemini_image_payload"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gemini/images_spec.rb","notes":"Endpoint routing on Vertex AI.","example":"uses Vertex model paths for Gemini image models"}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate and animate_later use Veo predictLongRunning, poll fetchPredictOperation and return inline or Cloud Storage output. Reference images take one image. Veo is absent from the bundled registry, so use a documented model with assume_model_exists: true. Omni generation is a separate row.","sources":["s614","s308","s309"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/videos.rb","method":"video_url, video_job_url, refresh_video_job, download_video"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"veo-3.1-fast-generate-001","spec":"spec/ruby_llm/providers/vertexai/videos_spec.rb","example":"generates a video through a Vertex prediction job","cassette":"spec/fixtures/vcr_cassettes/providers_vertexai_videos_generates_a_video_through_a_vertex_prediction_job.yml","notes":"Submission, polling and MP4 bytes in us-central1."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak calls generateContent on Gemini TTS models and Cloud Text-to-Speech text:synthesize for Chirp3-HD voices. Recordings cover the Chirp3-HD route only; Gemini TTS on Vertex AI is unrecorded and returns PCM.","sources":["s615","s607","s616"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"speech_url"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/text_to_speech.rb","method":"speech_url, render_speech_payload"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"audio_protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-11-03","model":"Chirp3-HD","spec":"spec/ruby_llm/protocols/vertexai/text_to_speech_spec.rb","example":"RubyLLM::Protocols::VertexAI::TextToSpeech live speaks in the language of the voice and the format asked for","cassette":"spec/fixtures/vcr_cassettes/protocols_vertexai_texttospeech_live_speaks_in_the_language_of_the_voice_and_the_format_asked_for.yml","notes":"text:synthesize returns Opus audio from the EU endpoint."}]},"transcription":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.transcribe uses generateContent with audio on Gemini models and audioTranscriptionConfig on the dedicated preview model. Returns a typed transcript.","sources":["s410","s593"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini.rb","method":"transcription_url"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/transcription.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-2.5-flash","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality vertexai/gemini-2.5-flash can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_vertexai_gemini-2_5-flash_can_transcribe_audio.yml","notes":"Transcript through generateContent."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe with gemini-3.5-transcribe-live-preview streams BidiGenerateContent partial and final text into TranscriptionChunk and returns a Transcription. It takes mono 16-bit PCM WAV and requires the global location. Google names the current model gemini-3.5-transcribe-live, which RubyLLM does not route to this protocol. Diarization and word timestamps raise on this route. This single operation is separate from live conversations.","sources":["s410","s215"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"transcription_protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/live_transcription.rb","method":"validate_transcription_request, websocket_service"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/live_transcription.rb"}],"verification":[{"status":"passed","recorded_on":"2026-09-17","model":"gemini-3.5-transcribe-live-preview","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription streams partial and final transcription through vertexai Live","cassette":"spec/fixtures/websocket_cassettes/transcription_google_vertexai.json","notes":"WebSocket setup, partial and final text, and generation complete."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"The dedicated transcription model takes speaker_names: [] to request detected speaker labels, with timestamps: :word for word offsets. Google marks diarization and word timestamps Experimental. Known-speaker identities are not supported. It requires the global location. The Live route does not diarize.","sources":["s410"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/transcription.rb","method":"render_transcription_payload, parse_transcription_segment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe-preview","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription transcribes audio with speaker labels and word timestamps through vertexai","cassette":"spec/fixtures/vcr_cassettes/transcription_transcribes_audio_with_speaker_labels_and_word_timestamps_through_vertexai.yml","notes":"audioTranscriptionConfig returns speaker labels and word offsets."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe-preview","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription distinguishes two speakers through vertexai dedicated transcription","cassette":"spec/fixtures/vcr_cassettes/transcription_distinguishes_two_speakers_through_vertexai_dedicated_transcription.yml","notes":"Two distinct speaker labels."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The audited generative endpoints have no standalone moderation operation. Google documents prompting Gemini as a filter, safety settings and Llama Guard settings on generation requests, none of which is a moderation API. Model Armor is a separate service.","sources":["s617","s618","s619"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"ocr":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.ocr reaches Mistral OCR (25.05) through the Vertex Mistral protocol's rawPredict route. The partner model must be enabled in Model Garden, and it answers in us-central1 and europe-west4, not in the global location, so vertexai_location must name one of those. DeepSeek-OCR had a scheduled retirement date of October 21, 2026, which has passed. Chat-prompted OCR on Gemini is not counted. Cloud Vision and Document AI are separate services.","sources":["s620","s621","s585"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/mistral.rb","method":"ocr_url"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/ocr.rb","method":"render_ocr_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"mistral-ocr-2505","spec":"spec/ruby_llm/ocr_spec.rb","example":"RubyLLM::OCR basic functionality vertexai/mistral-ocr-2505 extracts markdown from a PDF","cassette":"spec/fixtures/vcr_cassettes/ocr_basic_functionality_vertexai_mistral-ocr-2505_extracts_markdown_from_a_pdf.yml","notes":"us-central1 read the sample PDF as one page of markdown and reported pages_processed 1."}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"Text embedding models use predict. gemini-embedding-2 uses embedContent and takes one input per request, so several texts are separate requests.","sources":["s622","s318"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/embeddings.rb","method":"embedding_url"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/embed_content.rb","method":"embedding_url, render_embedding"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"text-embedding-004","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality vertexai/text-embedding-004 can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_vertexai_text-embedding-004_can_handle_a_single_text.yml","notes":"predict returns one vector."}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"embed(..., with:) sends text and image, audio, video or PDF parts to gemini-embedding-2 through embedContent and returns one combined vector. The older multimodalembedding endpoint, with separate per-modality vectors, is outside this integration. Text plus image is recorded; other modalities are not.","sources":["s318"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/embed_content.rb","method":"render_embedding_payload, parse_embedding_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-embedding-2","spec":"spec/ruby_llm/providers/vertexai/embed_content_spec.rb","example":"embeds text and an image through Vertex Gemini Embedding 2","cassette":"spec/fixtures/vcr_cassettes/providers_vertexai_embedcontent_embeds_text_and_an_image_through_vertex_gemini_embedding_2.yml","notes":"embedContent with a text and an image part."}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.rerank uses the Discovery Engine ranking API with the same Google credentials, the default ranking config unless overridden, and stable record IDs that recover the original text. It supports top_n and provider options. The Discovery Engine API must be enabled. The ranker is absent from the bundled registry, so use assume_model_exists: true.","sources":["s398","s399"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"ranking_config, ranking_connection"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/ranking.rb","method":"rerank_url, render_rerank_payload, parse_rerank_response"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/ranking_spec.rb","notes":"Request, endpoint and result mapping.","example":"uses Discovery Engine with existing credentials and correlates records that omit their original text"},{"status":"unavailable","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/ranking_spec.rb","notes":"The live example has no committed cassette.","example":"ranks real documents through the configured Discovery Engine service","model":"semantic-ranker-default-004"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload writes Cloud Storage objects with google-cloud-storage and returns gs:// references for Gemini requests and batches. This is Cloud Storage, not a Gemini Files API. Google-managed Claude does not support the Files API, and its file references are rejected.","sources":["s623","s592","s591"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/files.rb","method":"upload, find, download"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic.rb","method":"provider_file_references?"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/files_spec.rb","notes":"Upload, lookup and download against a stubbed bucket."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download and UploadedFile read Cloud Storage objects by gs:// URI.","sources":["s623"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/files.rb","method":"download, find"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/files_spec.rb","notes":"Upload, lookup and download against a stubbed bucket."}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch submits batchPredictionJobs through Cloud Storage for Gemini, Claude and MaaS chat models. Claude needs a regional location. Mistral models have no batch route. Requires google-cloud-storage and vertexai_batch_gcs_uri.","sources":["s624","s625","s626"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"batch_protocol_for, batch_protocol_for_model_path"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini/batches.rb","method":"vertex_batch_request"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/anthropic/batches.rb","method":"vertex_batch_request"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/chat_completions/batches.rb","method":"vertex_batch_request"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/vertexai/chat_completions/batches_spec.rb","notes":"JSONL rows and result parsing. Sibling specs cover Gemini and Claude.","example":"uses OpenAI JSONL rows for MaaS chat models"}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch supports text-embedding-004, text-embedding-005, text-multilingual-embedding-002, gemini-embedding-001 and gemini-embedding-2. Correlation keys restore shapes across shards. Google marks Gemini embedding batches Preview. Requires google-cloud-storage and vertexai_batch_gcs_uri.","sources":["s414","s624"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/embedding_prediction.rb","method":"MODELS, validate_batch_requests!"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/batch_prediction.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"batch_protocol_for"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/embedding_prediction_spec.rb","notes":"Request rows, correlation and result shapes.","example":"uploads legacy text instances with excluded correlation keys and job-wide dimensions"},{"status":"unavailable","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/vertexai/embedding_prediction_live_spec.rb","notes":"The live example needs a Cloud Storage prefix and has no committed cassette.","example":"completes text embeddings and restores shapes after reloading the prediction job","model":"text-embedding-004"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"Google-managed Claude does not offer the advisor tool; only Anthropic-managed Claude on claude.googleapis.com does. No Gemini equivalent is documented.","sources":["s591","s500","s603"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for the Perplexity Agent API.","sources":["s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"agent_skills":{"v1":"missing","v2":"missing","offered":true,"notes":"Managed Agents attach skills from Cloud Storage or Skill Registry. RubyLLM implements no agent or skill resources on Vertex AI; with_skills runs locally. Google-managed Claude does not offer Agent Skills.","sources":["s204","s591"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"async_research":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.research and research_later run single-turn hosted Deep Research jobs with polling, find by ID and cancellation. They require the global location. Usage comes from reported counters. Agent provisioning, public streaming and multi-turn research are not included.","sources":["s205","s203","s628"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/research.rb","method":"create_research_job, refresh_research_job, cancel_research_job"},{"version":"v2","path":"lib/ruby_llm/research_job.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deep-research-preview-04-2026","spec":"spec/ruby_llm/protocols/vertexai/research_spec.rb","example":"completes a short hosted research task through an actual public documentation MCP server","cassette":"spec/fixtures/vcr_cassettes/protocols_vertexai_research_completes_a_short_hosted_research_task_through_an_actual_public_documentation_mcp_server.yml","notes":"Submission, polling and completed report."},{"status":"passed","recorded_on":"2026-10-11","model":"deep-research-preview-04-2026","spec":"spec/ruby_llm/protocols/vertexai/research_spec.rb","example":"cancels an actual independently submitted hosted research task","cassette":"spec/fixtures/vcr_cassettes/protocols_vertexai_research_cancels_an_actual_independently_submitted_hosted_research_task.yml","notes":"POST to cancel; the provider confirms cancelled."}]},"background_responses":{"v1":"na","v2":"unknown","offered":null,"notes":"The Grok Responses page shows background: false in a sample response and documents no background mode. That field alone does not establish the feature. Checked the Grok Responses page and the Grok 4.7 model page. Interactions background execution is a separate row. A live check was not possible: xai/grok-4.3 returned 404, not found or no access, for this project.","sources":["s206","s629"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"xai/grok-4.3","notes":"POST endpoints/openapi/responses returned 404: the publisher model was not found or the project has no access."}]},"browser_use":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Google-managed Claude documents the browser use tool, and Gemini computer use documents browser environments. with_provider_tools accepts a raw declaration. The application executes the actions through tools registered with Tool.namespace.","sources":["s502","s591","s630"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The audited inventory documents no classification endpoint. Prompting a model to classify is not counted.","sources":["s627","s619"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"computer_use":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Gemini computer use is generally available, and Google-managed Claude documents the computer use tool. with_provider_tools accepts a raw declaration. The application executes the actions through tools registered with Tool.namespace.","sources":["s630","s503","s591","s631"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/tool.rb"}]},"context_editing":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Google-managed Claude documents context editing in beta. Clearing edits go through with_provider_options(context_management:) and with_headers for the beta header. with_compaction covers only the compaction edit.","sources":["s504","s603"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. Embedding task types are not contextual embeddings.","sources":["s622","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe selects the dedicated protocol for gemini-3.5-transcribe-preview, sends audioTranscriptionConfig and returns word offsets and speaker labels. It requires the global location. Google names the current model gemini-3.5-transcribe, which RubyLLM does not route to this protocol.","sources":["s410"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"transcription_protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/transcription.rb","method":"render_transcription_options, parse_transcription_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe-preview","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription transcribes audio with speaker labels and word timestamps through vertexai","cassette":"spec/fixtures/vcr_cassettes/transcription_transcribes_audio_with_speaker_labels_and_word_timestamps_through_vertexai.yml","notes":"audioTranscriptionConfig returns speaker labels and word offsets."},{"status":"passed","recorded_on":"2026-10-11","model":"gemini-3.5-transcribe-preview","spec":"spec/ruby_llm/transcription_google_spec.rb","example":"RubyLLM::Transcription distinguishes two speakers through vertexai dedicated transcription","cassette":"spec/fixtures/vcr_cassettes/transcription_distinguishes_two_speakers_through_vertexai_dedicated_transcription.yml","notes":"Two distinct speaker labels."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory.","sources":["s619","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Priority PayGo and Flex PayGo use the X-Vertex-AI-LLM-Shared-Request-Type header, available through with_headers. Claude fast mode is offered only on Anthropic-managed Claude.","sources":["s207","s632","s208"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_headers"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Cloud Storage lists and deletes objects. The public RubyLLM file API uploads, finds and downloads. The internal list_uris helper serves batch output only.","sources":["s623"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/files.rb","method":"list_uris"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"RAG Engine exposes corpus and file lifecycle. RubyLLM has no client for these resources. A raw retrieval tool can use an existing corpus.","sources":["s210"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"fill_in_middle":{"v1":"na","v2":"missing","offered":true,"notes":"Codestral 2 on Vertex AI documents fill-in-the-middle completion with a prompt and an optional suffix. The Vertex Mistral protocol sends only chat messages.","sources":["s633","s634"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai/mistral.rb","method":"completion_url"}]},"fine_tuning":{"v1":"missing","v2":"missing","offered":true,"notes":"Google offers supervised and reinforcement tuning jobs. RubyLLM has no tuning client.","sources":["s635","s636"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"google_maps":{"v1":"passthrough","v2":"native","offered":true,"notes":"The google_maps alias sends the googleMaps grounding tool on Gemini models.","sources":["s637","s202"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini.rb","method":"SERVER_TOOL_ALIASES"}]},"guardrail_configuration":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"safetySettings and modelArmorConfig on generateContent, and Llama Guard settings on Llama routes, go through with_provider_options. Separate Cloud security services are outside this row.","sources":["s638","s202","s639"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Vertex AI Interactions stores conversation state and continues it with previous_interaction_id. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s628","s640"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"image_batches":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Gemini image models support batch inference. RubyLLM.batch can stage a chat request with raw image response modalities and parse the attachments. There is no batched paint API.","sources":["s611","s624"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/gemini/batches.rb","method":"parse_vertex_batch_result"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/media.rb","method":"build_inline_attachment"}]},"interactions_api":{"v1":"missing","v2":"partial","offered":true,"notes":"The Research protocol reaches the Vertex Interactions endpoint for single-turn hosted agents. Chat cannot select Interactions on Vertex AI, although Google documents it for Gemini 3 and later models. Function tools, continuation and other agents are unreachable.","sources":["s628","s203","s640","s631"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/research.rb","method":"research_url"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Gemini generationConfig takes application/json without a schema, passed through with_provider_options. Chat Completions routes take response_format the same way.","sources":["s587","s202"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for Amazon Bedrock Mantle.","sources":["s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"manual_compaction":{"v1":"na","v2":"missing","offered":true,"notes":"Claude documents on-demand compaction on Google-managed Claude. Chat#compact needs a protocol with a compaction route, and only the Responses protocol has one.","sources":["s508"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb","method":"compaction_url"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render_compaction_payload"}]},"music_generation":{"v1":"missing","v2":"missing","offered":true,"notes":"Lyria 3 runs through Interactions and Lyria 2 through predict. RubyLLM has no music operation, and neither route is reachable.","sources":["s641","s642","s628"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Batch inference documents Gemini generation and embedding models, including image output. The TTS model page lists no batch option, and Veo and Omni pages mark batch inference not supported. Image and embedding batches have their own rows.","sources":["s624","s607","s614","s643"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"omni_video_generation":{"v1":"missing","v2":"missing","offered":true,"notes":"Gemini Omni 1.2 Flash generates video through Interactions and is generally available. RubyLLM implements Veo prediction jobs and has no Interactions route for it.","sources":["s643","s308","s631"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/videos.rb"}]},"partner_model_protocols":{"v1":"missing","v2":"native","offered":true,"notes":"protocol_for selects Claude rawPredict, Mistral rawPredict, or Chat Completions for publisher-prefixed MaaS models such as Grok, Llama, Qwen, GLM and DeepSeek. The Grok Responses route and Mistral OCR are separate rows.","sources":["s585","s586","s644","s634"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/vertexai_spec.rb","notes":"Routing for each publisher family.","example":"routes publisher-prefixed MaaS models to the Chat Completions protocol"}]},"prefix_completion":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Claude models before 4.6 accept a final assistant prefill: add the assistant message, then complete. Claude 4.6 and later reject it. The partner and open model pages document no prefix option.","sources":["s188","s591"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"Google-managed Claude does not offer programmatic tool calling, and no Gemini equivalent is documented.","sources":["s591","s510","s202"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching sets Claude TTL and automatic caching, and CachedContent attaches a Gemini cache resource.","sources":["s214","s605","s37"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_options"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/caches.rb","method":"render_cache_ttl"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"Claude server-side fallback is in beta only on the Claude API and Anthropic-managed Claude. No Gemini equivalent is documented.","sources":["s505","s202"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Vertex AI documents realtime conversations through the Live API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s215"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. Context caching stores prompt prefixes, not responses.","sources":["s214","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"responses":{"v1":"missing","v2":"missing","offered":true,"notes":"Google offers Grok through the Responses API at endpoints/openapi/responses, without stored state. RubyLLM Vertex AI registers no Responses protocol and routes Grok to Chat Completions.","sources":["s206"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb","method":"protocol_for"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Grok Responses page documents no batches, and the Grok 4.7 model page marks batch predictions not supported.","sources":["s206","s629"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. Web grounding is a tool on generation requests, and Vertex AI Search is a separate product.","sources":["s597","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory.","sources":["s203","s206"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. Image models generate images directly, and the Interactions tool list has no image generation tool.","sources":["s203","s609"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"server_tool_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_tools(defer: true) marks tools defer_loading and adds the tool search tool on Claude routes. Claude documents the tool search tool on Google-managed Claude. Gemini routes send deferred tools as ordinary tools.","sources":["s40","s591"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"apply_tool_search"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"tool_search?"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_deferred_tools_spec.rb","notes":"Claude on Vertex AI renders defer_loading. No recording covers this route.","example":"does not gate deferral on the model catalog"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. The Grok pages on Vertex AI document no X search tool.","sources":["s206","s629"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"shell_computer_tools":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Google-managed Claude documents the bash, text editor and computer tools, and Gemini documents computer use. Raw declarations go through with_provider_tools and the application executes them. This does not imply a hosted shell.","sources":["s511","s512","s503","s630"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. Lyria generates music.","sources":["s642","s619"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"streaming_speech":{"v1":"missing","v2":"missing","offered":true,"notes":"Gemini TTS streams audio through streamGenerateContent. RubyLLM.speak with a block raises on Vertex AI, because neither the Gemini nor the Cloud Text-to-Speech protocol implements streaming speech.","sources":["s615","s607"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"stream_speech"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/text_to_speech.rb"}]},"task_budgets":{"v1":"unknown","v2":"passthrough","offered":true,"notes":"Claude documents task budgets in beta on Google-managed Claude. The output_config setting and beta header go through with_provider_options and with_headers. RubyLLM has no named option.","sources":["s217","s603"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory.","sources":["s619","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"RAG Engine exposes corpus and file lifecycle. RubyLLM has no client for these resources. A raw retrieval tool can use an existing corpus.","sources":["s210"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Veo and Omni model pages mark batch inference not supported. Long-running single jobs are not batches.","sources":["s614","s643"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"video_editing":{"v1":"missing","v2":"missing","offered":true,"notes":"Google documents video editing on Gemini Omni through Interactions. RubyLLM implements Veo prediction jobs and has no Interactions route for it.","sources":["s213","s643"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/videos.rb"}]},"video_extension":{"v1":"missing","v2":"native","offered":true,"notes":"animate(extend:) sends a gs:// video or inline bytes to Veo predictLongRunning. Veo is absent from the bundled registry, so use a documented model with assume_model_exists: true. Inline bytes have unit coverage only. Omni extension through Interactions is not reachable.","sources":["s212","s614"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/vertexai/videos.rb","method":"render_video_extension"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/videos.rb","method":"render_video_extension_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"veo-3.1-fast-generate-001","spec":"spec/ruby_llm/providers/vertexai/videos_spec.rb","example":"extends a Cloud Storage video through a Vertex prediction job","cassette":"spec/fixtures/vcr_cassettes/providers_vertexai_videos_extends_a_cloud_storage_video_through_a_vertex_prediction_job.yml","notes":"gs:// sample extended in us-central1."}]},"voice_cloning":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Gemini TTS voice replication is in Preview. A replicated voice identifier goes in speechConfig through provider_options on speak. RubyLLM has no replication lifecycle, and the Chirp3-HD protocol rejects voice clone fields.","sources":["s645","s202"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/vertexai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gemini/speech.rb","method":"render_speech_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/vertexai/text_to_speech.rb","method":"RESERVED_OPTIONS"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory.","sources":["s619","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"Not documented in the audited inventory. URL context is a tool on generation requests.","sources":["s602","s627"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The Grok Responses page documents no WebSocket transport. Live API WebSockets are the realtime row.","sources":["s206"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/vertexai.rb"}]}},"gaps":[{"notes":"Remote MCP tools: Only hosted Deep Research connects a remote MCP server. Chat has no route, and the recorded research response omits MCP call and result steps.","sources":["s205","s203","s591"]},{"notes":"Interactions API: Only single-turn hosted research reaches the endpoint. Chat cannot select Interactions, so model conversations, function tools and other agents are unreachable.","sources":["s628","s203"]},{"notes":"Omni video generation, Video editing: Gemini Omni runs through Interactions, which RubyLLM does not route for video.","sources":["s643","s213"]},{"notes":"Fill-in-the-middle completion: Codestral 2 documents prompt and suffix completion. The Vertex Mistral protocol sends only chat messages.","sources":["s633"]},{"notes":"Manual context compaction: Claude documents on-demand compaction on Google-managed Claude. Chat#compact supports only the Responses protocol.","sources":["s508"]},{"notes":"Responses API: Grok Responses on Vertex AI has no route. Grok runs through Chat Completions.","sources":["s206"]},{"notes":"Music generation: Lyria runs through Interactions and predict. RubyLLM has no music operation.","sources":["s641","s628"]},{"notes":"Streaming speech generation: Gemini TTS streams through streamGenerateContent. speak with a block raises on Vertex AI.","sources":["s615"]},{"notes":"Agent skills: Managed Agents attach skills. RubyLLM has no agent or skill resources on Vertex AI.","sources":["s204"]},{"notes":"File listing and deletion: Cloud Storage offers both. The public file API uploads, finds and downloads.","sources":["s623"]},{"notes":"File search store management, Vector store management: RubyLLM has no RAG Engine corpus client.","sources":["s210"]},{"notes":"Fine-tuning: RubyLLM has no tuning job client.","sources":["s635"]},{"notes":"Token counting: chat.count_tokens raises on Claude routes, although Google documents a count-tokens endpoint for them.","sources":["s606"]},{"notes":"Dedicated transcription API, Streaming transcription: RubyLLM routes only the preview model ids to these protocols, while Google names gemini-3.5-transcribe and gemini-3.5-transcribe-live as the models.","sources":["s410"]},{"notes":"Web search tool: The Claude alias sends web_search_20260318, while Google-managed Claude documents only the basic tool.","sources":["s493","s601"]},{"notes":"Image input, PDF input: URL attachments on Claude routes render as URL sources, which Google-managed Claude does not support.","sources":["s591"]},{"notes":"Computer use, Browser use, Shell and computer tools, Context editing, Task budgets, Fast inference, JSON object mode, Guardrail configuration, Image batches, Prefix completion, Voice cloning: These work only through raw options.","sources":["s630","s502","s511","s504","s217","s207","s587","s638","s611","s188","s645"]}]},{"id":"bedrock","name":"Amazon Bedrock","notes":"Covers Bedrock foundation-model inference on bedrock-runtime and bedrock-mantle, plus documented inference resources and controls. Features may require an explicit protocol, model, region or IAM permission. AWS Data Automation, AgentCore, managed agents and general AWS services are outside scope.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.chat reaches every route the provider declares. Claude uses the Messages API through InvokeModel, or the Mantle Messages API for bare anthropic. ids. Nova, Llama and other versioned models use Converse. Mantle catalog models use Mantle Responses or Chat Completions.","sources":["s646","s647","s648","s649"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"protocol_for, chat_protocol_for, mantle_protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"completion_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality bedrock/amazon.nova-2-lite-v1:0 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_bedrock_amazon_nova-2-lite-v1_0_can_have_a_basic_conversation.yml","notes":"Converse route, response parsed into a Message."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"ConverseStream, InvokeModelWithResponseStream and Mantle SSE all stream through the block form of ask. The InvokeModel route decodes AWS event streams.","sources":["s647","s650","s651"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"stream_url, stream_events"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/streaming.rb","method":"extract_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/event_stream.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses bedrock/amazon.nova-2-lite-v1:0 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_bedrock_amazon_nova-2-lite-v1_0_supports_streaming_responses.yml","notes":"Recorded against converse-stream."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema renders output_config.format on Claude routes, outputConfig.textFormat on Converse and response_format or text.format on Mantle. GPT models on Converse also need additionalModelRequestFields.text.format.strict, which needs raw options.","sources":["s652","s653","s654"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"build_output_config"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_output_config"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat with schema with bedrock/claude-haiku-4-5 accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_bedrock_claude-haiku-4-5_accepts_a_json_schema_and_returns_structured_output.yml","notes":"InvokeModel route."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as Converse image blocks or Claude image blocks. Remote images on the InvokeModel route are sent as bytes, because Bedrock does not fetch URLs.","sources":["s655","s649","s656"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/media.rb","method":"render_image_attachment"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"inline_url_sources"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-sonnet-4-5-20250929-v1:0","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models bedrock/claude-sonnet-4-5 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_bedrock_claude-sonnet-4-5_can_understand_local_images.yml","notes":"InvokeModel route."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDF attachments work on Converse document blocks and on Claude document blocks. Claude on Bedrock takes PDFs inline or through Converse, not through a Files API.","sources":["s655","s656"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/media.rb","method":"render_document_attachment"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"provider_file_references?"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-sonnet-4-5-20250929-v1:0","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models bedrock/claude-sonnet-4-5 understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_bedrock_claude-sonnet-4-5_understands_pdfs.yml","notes":"InvokeModel route."}]},"audio_input":{"v1":"passthrough","v2":"native","offered":true,"notes":"Audio attachments render as Converse audio blocks, or as input_audio on Mantle Chat Completions for Voxtral. Claude routes do not take audio.","sources":["s655","s657","s658"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb","method":"render_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/media.rb","method":"render_audio_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_audio"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral.voxtral-mini-3b-2507","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models bedrock/mistral.voxtral-mini-3b-2507 can understand audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_bedrock_mistral_voxtral-mini-3b-2507_can_understand_audio.yml","notes":"Mantle Chat Completions with input_audio WAV. The model card does not describe this request shape."}]},"video_input":{"v1":"passthrough","v2":"native","offered":true,"notes":"Video attachments render as Converse video blocks for models that accept them, such as Nova 2 Lite.","sources":["s655","s647"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb","method":"render_attachment"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/media.rb","method":"render_video_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models bedrock/amazon.nova-2-lite-v1:0 can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_bedrock_amazon_nova-2-lite-v1_0_can_understand_local_videos.yml","notes":"MP4 sent in a Converse video block."}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Message#thinking returns Claude thinking blocks, Converse reasoningContent, and Responses reasoning items, with signatures kept for replay.","sources":["s659","s660","s661","s655"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"parse_thinking, parse_thinking_blocks"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_thinking_content"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking bedrock/claude-haiku-4-5 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_bedrock_claude-haiku-4-5_returns_thinking_when_available.yml","notes":"Budgeted thinking on InvokeModel."},{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat Bedrock Nova 2 reasoning returns reasoning content blocks","cassette":"spec/fixtures/vcr_cassettes/chat_bedrock_nova_2_reasoning_returns_reasoning_content_blocks.yml","notes":"Converse reasoningContent."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking maps effort and budget to thinking and output_config on Claude, reasoningConfig on Nova, and the matching Converse fields on other families. Each model accepts its own subset, and Nova rejects a token budget.","sources":["s659","s660","s661"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/reasoning.rb","method":"render_reasoning_fields"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_thinking_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking bedrock/claude-haiku-4-5 preserves thinking signatures between turns when provided","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_bedrock_claude-haiku-4-5_preserves_thinking_signatures_between_turns_when_provided.yml","notes":"Thinking settings and replay on InvokeModel."}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_citations enables document citations. RubyLLM parses Converse citationsContent and Claude citation blocks into Citation values, including search results returned by tools.","sources":["s655","s35","s649"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"parse_text_and_citations, parse_citation"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_text_and_citations"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with bedrock/claude-haiku-4-5 cites PDF documents with page numbers","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_bedrock_claude-haiku-4-5_cites_pdf_documents_with_page_numbers.yml","notes":"Response carries page_location citations."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Application tools work on Converse toolConfig, Claude tools, and Mantle function tools.","sources":["s662","s647","s663"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"render_tool, render_tool_config"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"function_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling bedrock/amazon.nova-2-lite-v1:0 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_bedrock_amazon_nova-2-lite-v1_0_can_use_tools.yml","notes":"Converse tool round trip."},{"status":"passed","recorded_on":"2026-10-11","model":"openai.gpt-oss-20b","spec":"spec/ruby_llm/providers/bedrock/mantle_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle responses chat calls tools through the Responses surface","cassette":"spec/fixtures/vcr_cassettes/providers_bedrock_mantle_responses_chat_calls_tools_through_the_responses_surface.yml","notes":"Mantle Responses tool round trip."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"choice maps to auto, required and a named tool on every route. Converse has no none value, so RubyLLM raises for choice: :none there. Claude Opus 5.5, Fable 5.1 and Mythos 5.1 reject forced tool use.","sources":["s664","s663"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"render_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control bedrock/amazon.nova-2-lite-v1:0 respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_bedrock_amazon_nova-2-lite-v1_0_respects_specific_tool_choice.yml","notes":"Sends toolChoice.tool on Converse."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool calls in one response are parsed and answered in a single user turn on Converse, Claude and Mantle routes.","sources":["s647","s663"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/streaming.rb"},{"version":"v1","path":"lib/ruby_llm/providers/bedrock/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"parse_tool_calls, render_messages"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling bedrock/amazon.nova-2-lite-v1:0 can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_bedrock_amazon_nova-2-lite-v1_0_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Converse response with several toolUse blocks."}]},"parallel_tool_control":{"v1":"passthrough","v2":"native","offered":true,"notes":"calls: :one sends disable_parallel_tool_use on Claude routes and parallel_tool_calls on Mantle routes. Converse documents no such field, and render_tool_config leaves the setting out on that route.","sources":["s503","s664","s651"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parallel_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"render_tool_config"}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends the nova_grounding system tool on Converse for Nova 2 models and returns ServerToolCall and Citation values. AWS Web Search for GPT-5.4 and later runs on Mantle Responses under /openai/v1, in US Regions and GovCloud, and RubyLLM posts there with protocol: :mantle_responses.","sources":["s665","s666","s227","s654"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse.rb","method":"SERVER_TOOL_ALIASES, merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES, web_search_sources?"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"completion_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with bedrock/amazon.nova-2-lite-v1:0 grounds the answer with web citations","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_bedrock_amazon_nova-2-lite-v1_0_grounds_the_answer_with_web_citations.yml","notes":"nova_grounding call with citationsContent in the response."}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_fetch) sends AWS Web Search, whose fetch step opens a page, on GPT-5.4 and later. The route needs protocol: :mantle_responses, because RubyLLM selects Chat Completions for these models by default, and a US Region or GovCloud. Fetching live pages needs the IAM permission bedrock-websearch:ExternalWebAccess, which AmazonBedrockFullAccess does not grant; an identity without it gets 401 access_denied unless it passes external_web_access: false, which limits Fetch to pages in the Bedrock cache. Nova grounding has no separate fetch operation.","sources":["s227","s654","s667","s651"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"SERVER_TOOL_ALIASES, completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle.rb","method":"RESPONSES_MODELS"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"mantle_api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"openai.gpt-5.6-luna","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat bedrock mantle web fetch opens a public page","cassette":"spec/fixtures/vcr_cassettes/chat_bedrock_mantle_web_fetch_opens_a_public_page.yml","notes":"Mantle Responses in us-west-2 with external_web_access: false returned a web_search_call whose open_page action read the requested URL."}]},"server_code_execution":{"v1":"na","v2":"native","offered":true,"notes":"with_provider_tools(:code_execution) adds Nova 2's nova_code_interpreter system tool on Converse, in IAD, PDX and NRT. It needs the bedrock:InvokeTool permission, which the default Bedrock API key role lacks. The code and its output come back as ServerToolCall values. Later turns rebuild from the local transcript as the answer text; the code and output are not replayed, the same as Nova grounding. Claude on Bedrock does not support code execution. AWS mentions a code interpreter for Mantle Responses without documenting the tool, and no recording covers it.","sources":["s666","s646","s649"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/converse.rb","method":"SERVER_TOOL_ALIASES, merge_server_tool_entries"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"parse_server_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat code execution with bedrock/amazon.nova-2-lite-v1:0 runs code server-side, keeps the result, and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_code_execution_with_bedrock_amazon_nova-2-lite-v1_0_runs_code_server-side_keeps_the_result_and_continues_from_the_local_transcript.yml","notes":"toolUse and toolResult blocks of nova_code_interpreter, the right product, and a second turn from local history."},{"status":"passed","recorded_on":"2026-11-10","model":"amazon.nova-2-lite-v1:0","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat code execution with bedrock/amazon.nova-2-lite-v1:0 streams a code interpreter run","cassette":"spec/fixtures/vcr_cassettes/chat_code_execution_with_bedrock_amazon_nova-2-lite-v1_0_streams_a_code_interpreter_run.yml","notes":"Streamed text matches the final message, with the tool call kept."}]},"server_file_search":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Knowledge Bases answer through RetrieveAndGenerate, which creates a provider-managed session for the conversation. RubyLLM does not integrate that lifecycle. Inference tools on Converse and Mantle document no file search tool.","sources":["s422","s220","s229"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(mcp: { connector_id: ... }) reaches Mantle Responses for Lambda function and AgentCore Gateway ARNs, and completed results replay statelessly. AWS examples set require_approval: never and do not use a public server_url. Claude on Bedrock does not support the MCP connector.","sources":["s229","s649"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"completion_url"}],"verification":[{"status":"unavailable","checked_on":"2026-09-07","model":"openai.gpt-oss-20b","notes":"No connector ARN was configured, so discovery and execution have no recording."},{"status":"unit","checked_on":"2026-09-07","spec":"spec/ruby_llm/providers/bedrock/mantle_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle request shape renders ARN-backed MCP connectors and replays their completed results"}]},"prompt_caching":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching places automatic cache points on Converse for Claude and Nova, and cache_control on Claude routes. Only eligible models and prompts cache, and models also apply implicit caching on their own.","sources":["s72"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb","method":"parse_completion_response"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"add_automatic_cache_points, converse_cache_block_for"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_control"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat prompt cache round-trip bedrock/claude-haiku-4-5 writes then reads the prompt cache","cassette":"spec/fixtures/vcr_cassettes/chat_prompt_cache_round-trip_bedrock_claude-haiku-4-5_writes_then_reads_the_prompt_cache.yml","notes":"Second call reads 7358 cached tokens on InvokeModel."}]},"cache_boundaries":{"v1":"passthrough","v2":"native","offered":true,"notes":"cache_until_here renders a Converse cachePoint or Claude cache_control with an optional TTL. Bedrock takes at most four points in one request, and Nova takes none in tools.","sources":["s72","s655"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock/chat.rb","method":"parse_completion_response"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"cache_boundary?, converse_cache_block_for"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"cache_boundary?"}],"verification":[{"status":"passed","recorded_on":"2026-10-27","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_cache_until_here_spec.rb","example":"RubyLLM::Chat Bedrock cache points caches a Claude turn that ends with Word and Excel documents","cassette":"spec/fixtures/vcr_cassettes/chat_bedrock_cache_points_caches_a_claude_turn_that_ends_with_word_and_excel_documents.yml","notes":"Cache point placed after the last block that accepts one."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Bedrock documents cache checkpoints in the prompt, with no separately created, found or deleted cache resource. Checkpoints count under cache boundaries.","sources":["s72","s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"compaction":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_compaction sends the compact_20260112 context_management edit on the Claude InvokeModel and Mantle routes, with its beta in anthropic_beta on InvokeModel. AWS lists Sonnet 4.6 and Opus 4.6 for this mechanism, and the Converse API supports neither compaction mechanism. No recording covers Bedrock.","sources":["s668","s496"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction, compaction_edit"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"render_invoke_body"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"count_tokens uses CountTokens on bedrock-runtime for Converse and InvokeModel bodies, addressed to the foundation model, and /anthropic/v1/messages/count_tokens on Mantle. Some Claude models support only the Mantle operation.","sources":["s73","s669"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"count_tokens_url, render_count_tokens_payload"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"count_tokens_url, post_count_tokens"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/anthropic.rb","method":"count_tokens_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with bedrock/claude-haiku-4-5 counts tokens with deferred tools declared","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_bedrock_claude-haiku-4-5_counts_tokens_with_deferred_tools_declared.yml","notes":"InvokeModel body wrapped for count-tokens."}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint supports Stable Diffusion 3.5 Large, Stable Image Core and Stable Image Ultra through InvokeModel. Size selects an aspect ratio, and each request returns one image. Nova Canvas reached end of life on September 30, 2026. Core and Ultra share the documented contract but have no recording.","sources":["s331","s332","s333","s670"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"image_protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/stability_images.rb","method":"paint, GENERATION_MODELS"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"stability.sd3-5-large-v1:0","spec":"spec/ruby_llm/protocols/invoke_model/stability_images_spec.rb","example":"RubyLLM::Protocols::InvokeModel::StabilityImages generates an image with Bedrock Stable Diffusion","cassette":"spec/fixtures/vcr_cassettes/protocols_invokemodel_stabilityimages_generates_an_image_with_bedrock_stable_diffusion.yml","notes":"Returned bytes decode to an image."}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint with a source image supports Stable Diffusion 3.5 image-to-image and Stable Image Inpaint with an explicit mask or the source alpha channel. Other Stability edit services, such as outpainting, upscaling, background removal and search and replace, raise an unsupported model error.","sources":["s331","s334","s335"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/stability_images.rb","method":"INPAINT_MODEL"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"image_protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"us.stability.stable-image-inpaint-v1:0","spec":"spec/ruby_llm/protocols/invoke_model/stability_images_spec.rb","example":"RubyLLM::Protocols::InvokeModel::StabilityImages inpaints a local image through a Bedrock inference profile","cassette":"spec/fixtures/vcr_cassettes/protocols_invokemodel_stabilityimages_inpaints_a_local_image_through_a_bedrock_inference_profile.yml","notes":"Source and mask through the inference profile."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate and animate_later use StartAsyncInvoke for Luma Ray 2 with one or two keyframe images. They require bedrock_video_s3_uri, and the completed MP4 downloads from the returned S3 prefix. Nova Reel reached end of life on September 30, 2026, and RubyLLM has no adapter for it. No completed generation has a recording.","sources":["s400","s671","s401","s402","s672"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"video_protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/async_videos.rb","method":"render_video_payload, refresh_video_job"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/files.rb","method":"list_uris, download"}],"verification":[{"status":"unavailable","checked_on":"2026-09-07","model":"luma.ray-v2:0","spec":"spec/ruby_llm/protocols/bedrock/async_videos_spec.rb","notes":"No job was submitted and no S3 object was written."},{"status":"unit","checked_on":"2026-09-07","spec":"spec/ruby_llm/protocols/bedrock/async_videos_spec.rb","example":"signs one submission, polls its ARN, and downloads the single MP4 under the returned output prefix","notes":"Covers the documented request and result lifecycle."}]},"speech":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Speech output on Bedrock uses Nova 2 Sonic and Nova 2.5 Sonic through InvokeModelWithBidirectionalStream, a voice conversation endpoint rather than standalone text to speech. RubyLLM does not implement realtime conversations.","sources":["s673","s79","s674"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"transcription":{"v1":"passthrough","v2":"native","offered":true,"notes":"RubyLLM.transcribe sends mistral.voxtral-small-24b-2507 audio as input_audio on Mantle Chat Completions and returns a typed Transcription with token usage. AWS describes the model as speech to text without documenting a transcription endpoint or this request shape. Timestamps, speaker labels and non-text formats raise, and other models do not route here.","sources":["s230","s657","s220"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"voxtral_transcription?"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/voxtral.rb","method":"transcribe, render_transcription_payload, parse_transcription_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral.voxtral-small-24b-2507","spec":"spec/ruby_llm/providers/bedrock/mantle/voxtral_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle::Voxtral transcribes speech through Bedrock Voxtral with typed usage","cassette":"spec/fixtures/vcr_cassettes/providers_bedrock_mantle_voxtral_transcribes_speech_through_bedrock_voxtral_with_typed_usage.yml","notes":"Successful transcript through Mantle Chat Completions."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe with a block sends the complete file to Mantle Voxtral and yields transcript deltas, then one final transcript. This streams output from an uploaded file, not live microphone input. A stream without a terminal stop raises.","sources":["s230","s657"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/voxtral.rb","method":"stream_transcription"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral.voxtral-small-24b-2507","spec":"spec/ruby_llm/providers/bedrock/mantle/voxtral_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle::Voxtral streams a complete multi-utterance WAV transcript through Bedrock Voxtral with typed usage","cassette":"spec/fixtures/vcr_cassettes/providers_bedrock_mantle_voxtral_streams_a_complete_multi-utterance_wav_transcript_through_bedrock_voxtral_with_typed_usage.yml","notes":"Successful streamed transcript with usage."}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"Foundation-model APIs document no speaker labels. Bedrock Data Automation and Amazon Transcribe are separate services.","sources":["s222","s230"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"moderation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.moderate(provider: :bedrock) calls ApplyGuardrail on the configured guardrail ID and version, for text and PNG or JPEG attachments. flagged? reflects GUARDRAIL_INTERVENED. Scores, tokens and cost stay nil. It creates no guardrail and does not use the detect-only InvokeGuardrailChecks operation. No recording covers it, because it needs an existing guardrail.","sources":["s81","s429","s675"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"guardrail_url, model_required?"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/guardrails.rb","method":"moderate"}],"verification":[{"status":"unavailable","checked_on":"2026-09-07","spec":"spec/ruby_llm/protocols/bedrock/guardrails_spec.rb","example":"applies an explicitly selected existing guardrail and reports its actual verdict","notes":"No guardrail ID and version were configured, so the live example has no cassette."},{"status":"unit","checked_on":"2026-09-07","spec":"spec/ruby_llm/protocols/bedrock/guardrails_spec.rb","notes":"Covers text and image verdicts, policy records and usage without a model."}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no OCR operation on foundation models. Bedrock Data Automation and Textract are separate services.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.embed routes Titan Text, Titan Multimodal, Cohere Embed and Nova Multimodal Embeddings models through InvokeModel, for text. Other embedding models on the platform, such as TwelveLabs Marengo, raise an unsupported model error.","sources":["s74","s677","s678","s333"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"embedding_protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/titan_text_embeddings.rb","method":"embed"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/cohere_embeddings.rb","method":"embed"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/nova_embeddings.rb","method":"embed"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"amazon.titan-embed-text-v2:0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality bedrock/amazon.titan-embed-text-v2:0 can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_bedrock_amazon_titan-embed-text-v2_0_can_handle_a_single_text.yml","notes":"Titan Text through InvokeModel."},{"status":"passed","recorded_on":"2026-10-11","model":"us.cohere.embed-v4:0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding Bedrock embedding dialects bedrock/us.cohere.embed-v4:0 embeds a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_bedrock_embedding_dialects_bedrock_us_cohere_embed-v4_0_embeds_a_single_text.yml","notes":"Cohere dialect through InvokeModel."}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"embed(with:) works on the three Bedrock multimodal families. Titan Multimodal takes one image alongside text. Cohere Embed v4 takes text and images as one input. Nova Multimodal takes exactly one image, audio or video per request, so text alongside an attachment raises ArgumentError; video uses AUDIO_VIDEO_COMBINED. Cohere Embed v3 on Bedrock refuses the inputs field and returns the provider's error.","sources":["s679","s678","s677"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/titan_multimodal_embeddings.rb","method":"embedding_image, render_embedding_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/cohere_embeddings.rb","method":"media_embedding_inputs"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/nova_embeddings.rb","method":"embedding_attachment, render_embedding_media"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"amazon.titan-embed-image-v1","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings bedrock/amazon.titan-embed-image-v1 embeds an image alongside text","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_bedrock_amazon_titan-embed-image-v1_embeds_an_image_alongside_text.yml","notes":"inputImage with inputText, 256 dimensions."},{"status":"passed","recorded_on":"2026-11-10","model":"us.cohere.embed-v4:0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings bedrock/us.cohere.embed-v4:0 embeds an image alongside text","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_bedrock_us_cohere_embed-v4_0_embeds_an_image_alongside_text.yml","notes":"Text and image in one inputs entry."},{"status":"passed","recorded_on":"2026-11-10","model":"amazon.nova-2-multimodal-embeddings-v1:0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings bedrock/amazon.nova-2-multimodal-embeddings-v1:0 embeds ruby_clip.mp4","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_bedrock_amazon_nova-2-multimodal-embeddings-v1_0_embeds_ruby_clip_mp4.yml","notes":"Video; image and audio are recorded alongside it."}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.rerank calls agent-runtime /rerank with Amazon Rerank or Cohere Rerank model ARNs, follows nextToken pages, and returns original indices with scores. Usage and cost stay unknown because the API does not return them. Cohere Rerank has no separate recording.","sources":["s76","s680"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"agent_api_base, agent_connection, rerank_model_arn"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/rerank.rb","method":"rerank, rerank_pages"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"amazon.rerank-v1:0","spec":"spec/ruby_llm/protocols/bedrock/rerank_spec.rb","example":"RubyLLM::Protocols::Bedrock::Rerank reranks actual documents with Amazon Rerank and leaves unreported usage unknown","cassette":"spec/fixtures/vcr_cassettes/protocols_bedrock_rerank_reranks_actual_documents_with_amazon_rerank_and_leaves_unreported_usage_unknown.yml","notes":"Top document index and positive score. Usage stays nil."},{"status":"unit","checked_on":"2026-09-07","spec":"spec/ruby_llm/protocols/bedrock/rerank_spec.rb","notes":"Covers pagination, repeated cursors, index bounds and JSON sources."}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload stores the file in S3 under bedrock_batch_s3_uri or an explicit s3:// uri, and needs the aws-sdk-s3 gem. Converse attaches S3 references for Nova models, and Claude routes send content inline. This is S3 storage, not a Bedrock files endpoint.","sources":["s655","s405","s224"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/files.rb","method":"upload"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"provider_file_references?"}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"Downloads and metadata lookups read s3:// objects through the same file API, using any credentials the bucket admits.","sources":["s224","s405"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/files.rb","method":"download, find"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch submits Model Invocation Jobs with S3 input and output. Claude requests use the InvokeModel format, which carries tools. Other models use the Converse format, where RubyLLM rejects tools and structured output. Bedrock also needs bedrock_batch_role_arn. The OpenAI Batch API on bedrock-runtime is not integrated. No recording covers these jobs.","sources":["s82","s405","s404","s406","s681"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/batches.rb","method":"create_batch, find_batch, cancel_batch, batch_results"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/batches.rb","method":"validate_bedrock_batch_requests!"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic/batches.rb","method":"bedrock_model_input"}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch accepts text embeddings for Titan Text Embeddings V2 and Titan Multimodal Embeddings G1 through InvokeModel jobs, with submission, refresh, cancellation and typed Embedding results restored to their request order. An S3 prefix and a batch role are required, and other embedding models raise. No completed job has a recording.","sources":["s74","s404","s405","s406"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"embedding_batch_protocol, embedding_batch_protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/invoke_model/embedding_batches.rb","method":"bedrock_batch_jsonl, parse_bedrock_outputs"}],"verification":[{"status":"unit","checked_on":"2026-09-07","spec":"spec/ruby_llm/protocols/invoke_model/embedding_batches_spec.rb","notes":"Covers the batch lifecycle with stubs, including shuffled records, errors and missing usage."},{"status":"unavailable","checked_on":"2026-09-07","model":"amazon.titan-embed-text-v2:0","spec":"spec/ruby_llm/protocols/invoke_model/embedding_batches_spec.rb","example":"RubyLLM::Protocols::InvokeModel::EmbeddingBatches submits, reloads and cancels an embedding batch using explicitly configured S3 resources","notes":"No S3 prefix or batch role was configured, so no job was submitted."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The Claude advisor tool page does not list Amazon Bedrock, and Claude in Amazon Bedrock lists it as not supported.","sources":["s500","s649"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for the Perplexity Agent API.","sources":["s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"Claude in Amazon Bedrock lists Agent Skills as not supported. AgentCore and managed agents are separate products.","sources":["s682","s649"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no asynchronous research operation.","sources":["s220","s646"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"background_responses":{"v1":"missing","v2":"partial","offered":true,"notes":"Mantle Responses documents background execution, and bedrock-runtime rejects it. Raw provider options can send background on Mantle, but RubyLLM has no retrieval, polling or cancellation lifecycle. Normal chat requests send store: false and replay local history.","sources":["s651","s683"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"store"}]},"browser_use":{"v1":"na","v2":"passthrough","offered":true,"notes":"Claude documents browser_toolset_20260801 on Bedrock through InvokeModel for Fable 5, Opus 4.8 and later models. with_provider_tools takes the raw declaration, and tools with namespace \"browser\" run the members locally. browser_state result blocks are not represented, so tab-management workflows are not covered.","sources":["s502","s656","s684"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no classification operation.","sources":["s220","s646"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"computer_use":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Bedrock documents computer_toolset_20260801 on newer Claude models and Anthropic-defined computer tools on others. with_provider_tools takes the raw declaration, and tools with namespace \"computer\" run the members locally with text and image results. Your application owns the environment. No recording covers Bedrock.","sources":["s684","s503","s656"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"}]},"context_editing":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Claude documents context editing on Bedrock and its InvokeModel route. It needs context_management through provider_options and the beta through with_headers, which the InvokeModel route moves into anthropic_beta.","sources":["s504"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"render_invoke_body"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no contextual embeddings operation.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"Voxtral is served through model inference, and the inventory lists no transcription endpoint. Data Automation and Amazon Transcribe are separate services.","sources":["s220","s230","s657"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no dubbing operation.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Converse performanceConfig.latency optimized is a preview feature for Nova Pro, Llama 3.1 and Claude 3.5 Haiku, sent through provider_options. Priority service tier is also a raw option. InvokeModel takes these settings through its own request fields, which with_headers can set. Claude fast mode does not list Bedrock.","sources":["s223","s685","s686","s650"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"S3 offers object listing and deletion. The public file API uploads, finds and downloads, and its internal listing serves batch output only.","sources":["s224"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/bedrock/files.rb","method":"list_uris"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"list_file_uris"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Knowledge Bases expose managed retrieval storage and lifecycle operations. RubyLLM has no Knowledge Base client.","sources":["s225","s220"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no fill-in-the-middle operation.","sources":["s220","s222"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"fine_tuning":{"v1":"missing","v2":"missing","offered":true,"notes":"Bedrock offers model customization jobs and OpenAI-compatible fine-tuning. RubyLLM does not manage either.","sources":["s83","s687"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for Google Maps grounding.","sources":["s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"guardrail_configuration":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Converse takes guardrailConfig through provider_options. InvokeModel takes guardrail request fields that with_headers can set. RubyLLM has no shared setting and maps guardrail_intervened to :content_filter.","sources":["s70","s688","s650"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"FINISH_REASONS"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Responses stores conversations for 30 days by default and continues them with previous_response_id. RubyLLM sends store: false and keeps conversation state in the application.","sources":["s651"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"store"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch model table lists no image generation model.","sources":["s404","s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for the Google Interactions API.","sources":["s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"Bedrock documents JSON Schema output, with no schema-free JSON object mode.","sources":["s652","s657"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"mantle_protocols":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM signs for bedrock-mantle and routes bare anthropic. ids to Messages, listed Gemma and GPT OSS models to Responses, and other Mantle catalog models to Chat Completions. OpenAI GPT models post under /openai/v1, as AWS requires, and everything else under /v1.","sources":["s683","s648","s651","s657","s689"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"mantle_protocol_for, mantle_api_base"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle.rb","method":"SIGNING_SERVICE, RESPONSES_MODELS"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/chat_completions.rb","method":"completion_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai.gpt-oss-20b","spec":"spec/ruby_llm/providers/bedrock/mantle_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle signed requests reaches the mantle endpoint with bedrock-mantle credentials","cassette":"spec/fixtures/vcr_cassettes/providers_bedrock_mantle_signed_requests_reaches_the_mantle_endpoint_with_bedrock-mantle_credentials.yml","notes":"Lists v1/models with a bedrock-mantle signature."}]},"manual_compaction":{"v1":"na","v2":"passthrough","offered":true,"notes":"Bedrock documents the compact-2026-09-04 top-level compaction parameter on InvokeModel for models such as Claude Opus 5.5 and Haiku 5.5, and not on Converse. Chat#compact has no render_compaction_payload on Bedrock protocols. The request and the adoption of the returned block need provider_options, with_headers and a before_request hook.","sources":["s668"],"code":[{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"compact"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb","method":"render_compaction_payload"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory and model catalog document no music generation.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch model table covers chat and embedding models, which have their own rows.","sources":["s404","s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for Google Gemini Omni.","sources":["s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"partner_model_protocols":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM selects Converse, Claude Messages on InvokeModel and Mantle, Responses, Chat Completions, and the Titan, Cohere, Nova, Stability and Luma InvokeModel formats for each model. Coverage is for these named formats, and other InvokeModel-only models fall back to Converse or raise.","sources":["s222","s646","s676"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"protocol_for, embedding_protocol_for, image_protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral.voxtral-mini-3b-2507","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models bedrock/mistral.voxtral-mini-3b-2507 can understand mp3 audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_bedrock_mistral_voxtral-mini-3b-2507_can_understand_mp3_audio.yml","notes":"Partner model through Mantle Chat Completions."}]},"prefix_completion":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Bedrock documents prefilling with a final assistant message on the Claude Messages format. RubyLLM has no prefix setter, so it takes an assistant message added before generating. Model support varies.","sources":["s226"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"render_messages"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The Claude page does not list Amazon Bedrock, and Claude in Amazon Bedrock lists it as not supported.","sources":["s510","s649"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching takes a TTL, and cache_until_here takes a per-message TTL, on Converse and Claude routes. TTLs depend on the model.","sources":["s72"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/converse/chat.rb","method":"prompt_cache_options, PROMPT_CACHE_OPTIONS"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"prompt_cache_options"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"Claude in Amazon Bedrock lists the fallbacks parameter as not supported. Inference requests document no other server-side model fallback.","sources":["s649","s656"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Nova 2 Sonic and Nova 2.5 Sonic use InvokeModelWithBidirectionalStream for realtime voice conversations. RubyLLM does not implement realtime conversations.","sources":["s673","s79","s674"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents prompt caching but no response cache.","sources":["s220","s72"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"Mantle Responses runs by default for listed Gemma and GPT OSS models and with protocol: :mantle_responses elsewhere, under /openai/v1 for OpenAI GPT models. The recommended bedrock-runtime /openai/v1 Responses path is not integrated.","sources":["s651","s690","s689"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb","method":"mantle_protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle/responses.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/mantle.rb","method":"RESPONSES_MODELS"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai.gpt-oss-20b","spec":"spec/ruby_llm/providers/bedrock/mantle_spec.rb","example":"RubyLLM::Providers::Bedrock::Mantle responses chat chats through the Responses surface","cassette":"spec/fixtures/vcr_cassettes/providers_bedrock_mantle_responses_chat_chats_through_the_responses_surface.yml","notes":"Successful v1/responses call."}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The OpenAI Batch API on Bedrock requires endpoint v1/chat/completions, so it documents no Responses batch.","sources":["s681"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no standalone web search operation. Knowledge Base Retrieve queries your own stores, and Web Search is a tool on inference requests.","sources":["s220","s227"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"Bedrock tool documentation does not list an apply patch tool.","sources":["s229","s227","s683"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Bedrock tool documentation does not list an image generation tool.","sources":["s229","s227","s683"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"server_tool_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_tools(defer: true) marks deferred tools and adds the BM25 search tool on the Claude InvokeModel route, with the tool-search beta in anthropic_beta. Claude documents this for InvokeModel and not for Converse.","sources":["s40"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock/anthropic.rb","method":"TOOL_SEARCH_BETA, beta_headers"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"apply_tool_search, NATIVE_TOOL_SEARCH"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"tool_search?"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with bedrock/claude-haiku-4-5 loads a deferred tool through tool search and calls it","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_bedrock_claude-haiku-4-5_loads_a_deferred_tool_through_tool_search_and_calls_it.yml","notes":"InvokeModel response with tool_search_tool_result blocks."},{"status":"passed","recorded_on":"2026-11-05","model":"us.anthropic.claude-haiku-4-5-20251001-v1:0","spec":"spec/ruby_llm/chat_tool_search_spec.rb","example":"RubyLLM::Chat deferred tools with bedrock/claude-haiku-4-5 loads a deferred tool while streaming","cassette":"spec/fixtures/vcr_cassettes/chat_deferred_tools_with_bedrock_claude-haiku-4-5_loads_a_deferred_tool_while_streaming.yml","notes":"Streamed through invoke-with-response-stream."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"Provider-specific row for X search. Bedrock tool documentation does not list it.","sources":["s229","s227"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"shell_computer_tools":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Claude on Bedrock documents client-executed bash, text editor, memory and computer tools. They need raw tool definitions through with_provider_tools and an application executor. A computer or browser toolset member can use a namespaced local tool.","sources":["s511","s649","s656","s684"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"CLIENT_TOOLSETS"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The inventory and model catalog document no sound effects generation.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"streaming_speech":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Speech output on Bedrock uses Nova 2 Sonic and Nova 2.5 Sonic through InvokeModelWithBidirectionalStream, a voice conversation endpoint rather than standalone text to speech. RubyLLM does not implement realtime conversations.","sources":["s673","s79","s674"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"task_budgets":{"v1":"unknown","v2":"passthrough","offered":true,"notes":"Regular Claude requests on Bedrock accept output_config.task_budget with task-budgets-2026-03-13 in anthropic_beta, both through with_provider_options. RubyLLM has no named option, and the budget is advisory. Claude's own task budgets page does not list Amazon Bedrock.","sources":["s217","s668","s659"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"global.anthropic.claude-haiku-5-5","notes":"A budget of 25000 returned a normal reply. A budget of 1 returned 400 asking for at least 1024, and leaving out the beta value returned 400: Extra inputs are not permitted, so Bedrock reads and validates the field."}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no text analysis operation.","sources":["s220","s646"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Knowledge Bases manage vector stores through their lifecycle operations. RubyLLM has no Knowledge Base client.","sources":["s225","s220"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/bedrock.rb"},{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch model table lists no video generation model.","sources":["s404","s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"Luma Ray 2 documents generation from text and keyframes, with no video editing input.","sources":["s400","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"Luma Ray 2 documents generation from text and keyframes, with no extension of an existing video.","sources":["s400","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The inventory and model catalog document no voice cloning.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The inventory and model catalog document no voice conversion.","sources":["s220","s676"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no standalone web fetch operation. Fetch runs inside the Web Search tool.","sources":["s220","s227"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The Responses documentation describes HTTP and SSE requests and no WebSocket transport. Bidirectional streaming on Bedrock is a separate Sonic operation.","sources":["s651","s220"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/bedrock.rb"}]}},"gaps":[{"notes":"Bedrock Mantle protocols, Responses API: The recommended bedrock-runtime /openai/v1 Responses and Chat Completions paths are not integrated.","sources":["s683","s651","s657","s689"]},{"notes":"Background Responses jobs: Mantle Responses documents background execution. Raw provider options can express the request field, but RubyLLM has no retrieval, polling or cancellation lifecycle.","sources":["s651"]},{"notes":"Image editing: Only Stable Diffusion 3.5 image-to-image and Stable Image Inpaint are integrated. Outpainting, upscaling, background removal, search and replace, and the other Stability edit services are not.","sources":["s334"]},{"notes":"Moderation: RubyLLM needs an existing guardrail and does not use the detect-only InvokeGuardrailChecks operation.","sources":["s81","s675"]},{"notes":"Chat batches: The Converse batch format rejects tools, as Bedrock does, and RubyLLM also rejects structured output there, which Bedrock accepts. The OpenAI Batch API on bedrock-runtime is not integrated.","sources":["s82","s681"]},{"notes":"Context compaction, Manual context compaction: with_compaction sends only the compact-2026-01-12 mechanism, listed for Sonnet 4.6 and Opus 4.6. The compact-2026-09-04 parameter needs raw options, because Chat#compact has no Bedrock implementation.","sources":["s668"]},{"notes":"Computer use, Browser use, Shell and computer tools, Context editing: These Claude features need provider-specific tool or context declarations and application-managed execution or state. Browser toolsets do not represent browser_state results.","sources":["s684","s502","s504","s656"]},{"notes":"Fast inference: Converse performanceConfig.latency and service tiers need raw options.","sources":["s223","s686"]},{"notes":"File listing and deletion: S3 offers object listing and deletion. The public file API only uploads, finds and downloads.","sources":["s224"]},{"notes":"File search store management, Vector store management: Knowledge Bases expose managed retrieval storage and lifecycle operations. RubyLLM has no Knowledge Base client.","sources":["s225"]},{"notes":"Fine-tuning: Bedrock offers model customization and OpenAI-compatible fine-tuning jobs. RubyLLM does not manage them.","sources":["s83","s687"]},{"notes":"Guardrail configuration: guardrailConfig and the InvokeModel guardrail fields need raw options, without a shared setting.","sources":["s70","s688"]},{"notes":"Prefix completion: A final assistant message can prefill on the Claude Messages format, without a shared prefix setter.","sources":["s226"]},{"notes":"Task budgets: Bedrock accepts output_config.task_budget with the beta value in anthropic_beta, through with_provider_options, with no named RubyLLM option.","sources":["s217","s668"]}]},{"id":"azure","name":"Azure OpenAI / Foundry","notes":"Covers Azure OpenAI v1 and the OpenAI-compatible routes in the Foundry REST reference, plus partner models documented on Foundry, including Cohere embeddings and reranking and Claude Messages. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products. Most media operations have Azure-specific endpoint handling with HTTP regression coverage but no live recording, and they require compatible deployments.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"chat uses Chat Completions or Responses on the deployment. Chat Completions is the default route; deployments named after gpt-5.4 or later, or gpt-6 and later, route to Responses. An explicit protocol: or azure_protocol overrides the routing.","sources":["s8","s691"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality azure/grok-4-1-fast-non-reasoning can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_azure_grok-4-1-fast-non-reasoning_can_have_a_basic_conversation.yml","notes":"HTTP 200 from /models/chat/completions."},{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.6-luna","spec":"spec/ruby_llm/providers/azure/responses_spec.rb","example":"RubyLLM::Providers::Azure::Responses a GPT-5.6 deployment routes to Responses without being asked","cassette":"spec/fixtures/vcr_cassettes/providers_azure_responses_a_gpt-5_6_deployment_routes_to_responses_without_being_asked.yml","notes":"The deployment name routes to /openai/v1/responses."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Both Chat Completions and Responses stream through server-sent events.","sources":["s8","s692"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses azure/grok-4-1-fast-non-reasoning supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_azure_grok-4-1-fast-non-reasoning_supports_streaming_responses.yml","notes":"Streamed Chat Completions recording."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends json_schema on Chat Completions and text.format on Responses. Support depends on the deployed model.","sources":["s693","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"schema_format"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat with schema with azure/grok-4-1-fast-non-reasoning allows removing schema mid-conversation","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_azure_grok-4-1-fast-non-reasoning_allows_removing_schema_mid-conversation.yml","notes":"Request sends response_format json_schema."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments work on both routes, as image_url parts or input_image parts. The Azure dialect keeps original image detail.","sources":["s694","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models azure/grok-4-1-fast-non-reasoning can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_azure_grok-4-1-fast-non-reasoning_can_understand_local_images.yml","notes":"Request sends an image_url part."}]},"pdf_input":{"v1":"missing","v2":"native","offered":true,"notes":"PDFs work on the Responses route as base64, uploaded file or URL. Azure Chat Completions attachments raise for PDFs in RubyLLM, so older deployment names need protocol: :responses. Azure rejects the user_data upload purpose, so uploads use assistants.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/media.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat document URL models azure/gpt-5.6-luna reads PDFs the provider fetches from their URL","cassette":"spec/fixtures/vcr_cassettes/chat_document_url_models_azure_gpt-5_6-luna_reads_pdfs_the_provider_fetches_from_their_url.yml","notes":"Responses request sends input_file with file_url."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Audio attachments render as input_audio on Chat Completions for audio-enabled deployments. The Responses route does not take audio attachments. No live recording covers this.","sources":["s695","s692"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/media.rb"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"Neither the Chat Completions nor the Responses content schema documents a video attachment part. Vision-enabled chats cover images.","sources":["s696","s692","s694"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Reasoning tokens are reported on both routes. Reasoning summaries and encrypted reasoning return on Responses, which replays them with store false.","sources":["s697","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_reasoning_summary"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5-nano","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking azure/gpt-5-nano returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_azure_gpt-5-nano_returns_thinking_when_available.yml","notes":"Request sends reasoning_effort; the spec checks reported thinking tokens."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking sends reasoning_effort on Chat Completions and reasoning.effort on Responses. Values vary by model.","sources":["s697","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5-nano","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking azure/gpt-5-nano returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_azure_gpt-5-nano_returns_thinking_when_available.yml","notes":"Request sends reasoning_effort; the spec checks reported thinking tokens."}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"The Responses route returns typed URL, file and container-file citations, in sync and streaming parsing. Azure documents URL citations for web search.","sources":["s10"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/citation.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"gpt-5-nano","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with azure/gpt-5-nano searches, returns the sources, and counts the searches Azure bills","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_azure_gpt-5-nano_searches_returns_the_sources_and_counts_the_searches_azure_bills.yml","notes":"Response holds web_search_call items with search and open_page actions, and url_citation annotations."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools work on both routes. Azure rejects function tools with reasoning on gpt-5.6-sol Chat Completions; recent GPT deployment names route to Responses.","sources":["s698","s697"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"render_tools"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling azure/grok-4-1-fast-non-reasoning can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_azure_grok-4-1-fast-non-reasoning_can_use_tools.yml","notes":"Tool round trip on Chat Completions."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"Tool choice maps to tool_choice on both routes.","sources":["s698","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control azure/grok-4-1-fast-non-reasoning respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_azure_grok-4-1-fast-non-reasoning_respects_specific_tool_choice.yml","notes":"Request sends a named tool_choice."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool calls in one response are executed on both routes. Support depends on the deployed model.","sources":["s698"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_function_calls"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling azure/grok-4-1-fast-non-reasoning can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_azure_grok-4-1-fast-non-reasoning_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Response carries several tool calls."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"calls: :one sends parallel_tool_calls false on both routes.","sources":["s696","s692"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control azure/grok-4-1-fast-non-reasoning respects calls: :one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_azure_grok-4-1-fast-non-reasoning_respects_calls_one_for_sequential_execution.yml","notes":"Request sends parallel_tool_calls false."}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) works on the Responses route, with options such as domain filters. RubyLLM asks for action sources and counts billed searches.","sources":["s10"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/responses_spec.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"gpt-5-nano","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with azure/gpt-5-nano searches, returns the sources, and counts the searches Azure bills","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_azure_gpt-5-nano_searches_returns_the_sources_and_counts_the_searches_azure_bills.yml","notes":"Response holds web_search_call items with search and open_page actions, and url_citation annotations."}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"Page opening and find-in-page are web_search actions on reasoning models, available through the web_search alias. Azure does not return their results through web_search_call.results.","sources":["s10"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"gpt-5-nano","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with azure/gpt-5-nano searches, returns the sources, and counts the searches Azure bills","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_azure_gpt-5-nano_searches_returns_the_sources_and_counts_the_searches_azure_bills.yml","notes":"Response holds web_search_call items with search and open_page actions, and url_citation annotations."}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"The code_execution alias sends code_interpreter with an automatic container on the Responses route. No Azure live recording covers it.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}]},"server_file_search":{"v1":"missing","v2":"native","offered":true,"notes":"The file_search alias works on the Responses route with a vector store prepared outside RubyLLM. Azure documents the tool in the Responses schema and vector stores reference. No Azure live recording covers it.","sources":["s696","s699"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"The mcp alias connects remote tools on the Responses route. Pending approvals appear among pending_questions, and approving or denying resumes with replayed history. Local tools with matching names never execute.","sources":["s8","s338"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/approvals.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb"},{"version":"v2","path":"lib/ruby_llm/tool_call.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/chat_methods.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/tool_call.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools azure executes a remote read-only tool after approval using stateless continuation","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_azure_executes_a_remote_read-only_tool_after_approval_using_stateless_continuation.yml","notes":"Public read-only Microsoft Learn MCP tool; continuation replays history with store false."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools azure denies a remote call without executing it","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_azure_denies_a_remote_call_without_executing_it.yml","notes":"Public read-only Microsoft Learn MCP tool; continuation replays history with store false."},{"status":"passed","recorded_on":"2026-11-06","model":"gpt-5.6-luna","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools azure streams a remote approval and its completed result","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_azure_streams_a_remote_approval_and_its_completed_result.yml","notes":"Public read-only Microsoft Learn MCP tool; continuation replays history with store false."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Automatic caching needs no option, and cached tokens are reported on both routes. with_caching sets options on supported models. Writes bill separately on GPT-5.6 and later.","sources":["s11"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"apply_prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"}]},"cache_boundaries":{"v1":"passthrough","v2":"native","offered":true,"notes":"Explicit breakpoints work on GPT-5.6 and later Standard deployments, on both routes. Azure rejects them on earlier models and does not support them on PTU-M deployments.","sources":["s11"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure/media.rb","method":"format_content"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"inject_cache_breakpoint"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"inject_cache_breakpoint"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"No separately created cache resource is documented. Azure caching uses automatic prefixes, keys and breakpoints, and Claude on Foundry uses request-level caching.","sources":["s11","s700","s701"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"compaction":{"v1":"missing","v2":"native","offered":true,"notes":"with_compaction sends context_management with a compact threshold on the Responses route and preserves the compaction item for later turns. The Chat Completions route has no compaction.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"apply_compaction"},{"version":"v2","path":"lib/ruby_llm/chat.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/chat_methods.rb"},{"version":"v2","path":"lib/ruby_llm/agent.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with azure","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_azure.yml","notes":"HTTP 200 from /openai/v1/responses/compact returning response.compaction, then a continued response."}]},"token_counting":{"v1":"na","v2":"missing","offered":true,"notes":"Claude on Foundry documents POST /v1/messages/count_tokens. RubyLLM Azure has no Messages route, and Azure::Responses does not include the input-token module. The Azure OpenAI inventories document no count endpoint.","sources":["s701","s696","s149"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/token_counting.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"image_generation":{"v1":"partial","v2":"native","offered":true,"notes":"paint sends Azure image requests and parses URL and base64 images. Resource and /openai/v1 bases use the v1 endpoints with api-version preview. Deployment bases keep their path, defaulting to 2025-04-01-preview. Requires a compatible deployment. No Azure live recording covers it. Partner image models with their own routes, such as FLUX, are not covered.","sources":["s149","s312","s313"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"image_editing":{"v1":"partial","v2":"native","offered":true,"notes":"paint with with: and mask: uses Azure multipart uploads, including for GPT image models, and accepts custom deployment names. Requires a compatible deployment. No Azure live recording covers it.","sources":["s149","s312","s313"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate uses the documented /openai/v1/video/generations/jobs create, poll and download flow. Width, height and duration go through provider_options. Reference images raise; the /videos route is not used. The retirement schedule shows 2026-10-15 for sora-2 version 2025-12-08 while the video guide and region tables still list it. No Azure live recording covers it.","sources":["s14","s149","s702"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/videos.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/videos_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/videos_spec.rb","notes":"Covers the jobs endpoint, payload, status parsing and failure reasons."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak posts to Azure audio/speech and returns the binary Speech result. Requires a compatible deployment. No Azure live recording covers it.","sources":["s149","s312","s342"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/audio.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"transcription":{"v1":"partial","v2":"native","offered":true,"notes":"transcribe sends multipart audio with language, prompt and format, and keeps the transcript and usage. Requires a compatible deployment. No Azure live recording covers it.","sources":["s149","s312","s703"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/audio.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe streams prerecorded audio through audio/transcriptions as server-sent events. whisper-1 does not stream. Realtime transcription sessions are separate. No Azure live recording covers it.","sources":["s149"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/audio.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"diarization":{"v1":"partial","v2":"native","offered":true,"notes":"The diarized_json parser keeps speaker labels on a gpt-4o-transcribe-diarize deployment. Custom deployment names pass format: \"diarized_json\". No Azure live recording covers it.","sources":["s149","s704"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"azure_media_url"},{"version":"v2","path":"lib/ruby_llm/providers/azure/audio.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions_operations_spec.rb","notes":"HTTP stubs cover Azure endpoints for images, speech and transcription."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"Not served on the audited Foundry resource: POST /openai/v1/moderations returned 404 on both its hosts and on its project endpoint, where chat completions answered, the same as a nonsense route, and none of the 492 models in GET /openai/v1/models is a moderation model. The Foundry REST reference, whose base is the project endpoint, still lists the path. Azure AI Content Safety is a separate product. RubyLLM.moderate raises RubyLLM::Error for Azure before a moderation request.","sources":["s705","s704","s155"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/moderation.rb","method":"render_moderation_payload"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"omni-moderation-latest","notes":"POST returned 404 Resource not found on /openai/v1/moderations, with and without api-version, and on /moderations, /openai/moderations, /openai/deployments/omni-moderation-latest/moderations and /models/moderations, as did a nonsense path. On the project endpoint /api/projects/PROJECT, /openai/v1/chat/completions returned 200 and /openai/v1/moderations returned the 404 of a nonsense path. text-moderation-latest returned 404 on /openai/v1/moderations. GET /openai/v1/models listed 492 models, none with moderation in its id."}]},"ocr":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.ocr reaches Foundry's Mistral document models through /providers/mistral/azure/ocr, with the deployment name as the model. Foundry also sells mistral-ocr-4-0 on the same route. Cohere documents Parse on Foundry, but RubyLLM has no route for it: Azure::Cohere leaves ocr_url abstract. Azure Document Intelligence is a separate product.","sources":["s704","s152","s156"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/azure/mistral.rb","method":"ocr_url"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/ocr.rb","method":"render_ocr_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"mistral-document-ai-2512","spec":"spec/ruby_llm/ocr_spec.rb","example":"RubyLLM::OCR basic functionality azure/mistral-document-ai-2512 extracts markdown from a PDF","cassette":"spec/fixtures/vcr_cassettes/ocr_basic_functionality_azure_mistral-document-ai-2512_extracts_markdown_from_a_pdf.yml","notes":"A pay-per-use GlobalStandard deployment read the sample PDF as one page of markdown and reported pages_processed 1."}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"embed uses the Azure embeddings route, and the Cohere v2 route for known Cohere deployment names. Other Cohere deployment names need azure_protocol: :cohere.","sources":["s706","s152","s704"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/cohere.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"Cohere-embed-v3-english","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality azure/Cohere-embed-v3-english can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_azure_cohere-embed-v3-english_can_handle_a_single_text.yml","notes":"HTTP 200 from /providers/cohere/v2/embed."}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"Cohere deployments on Foundry embed images through /providers/cohere/v2/embed. Embed v3 takes one image without text and v4 takes ordered mixed inputs. Only Cohere-embed-v3-english, Cohere-embed-v3-multilingual and embed-v-4-0 route automatically; other names, such as Embed v5, need azure_protocol: :cohere. Only v3 has a live recording.","sources":["s152","s396","s397","s704"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/cohere.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/embeddings.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"Cohere-embed-v3-english","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","example":"RubyLLM::Providers::Azure::Cohere embeds an image through the active Azure Cohere endpoint with billed usage","cassette":"spec/fixtures/vcr_cassettes/providers_azure_cohere_embeds_an_image_through_the_active_azure_cohere_endpoint_with_billed_usage.yml","notes":"HTTP 200 from /providers/cohere/v2/embed for one image."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","notes":"Covers URL normalization, authentication, image limits, routing and typed rerank results."}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"rerank uses the Azure Cohere resource or serverless endpoint and keeps original documents, positions and scores. Tokens and cost stay unknown when not reported. The live recording holds a 404, so no successful live rerank is recorded.","sources":["s151","s152","s704"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/cohere.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/rerank.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","notes":"Covers URL normalization, authentication, image limits, routing and typed rerank results."},{"status":"unavailable","checked_on":"2026-10-11","model":"Cohere-rerank-v4.0-fast","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","example":"RubyLLM::Providers::Azure::Cohere reranks documents through the configured Azure Cohere deployment","cassette":"spec/fixtures/vcr_cassettes/providers_azure_cohere_reranks_documents_through_the_configured_azure_cohere_deployment.yml","notes":"The recording holds HTTP 404. No successful rerank is recorded."}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"upload uses /openai/v1/files. Responses attachment uploads use purpose assistants.","sources":["s435","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/azure/files.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb","method":"provider_file_upload_options"}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"download uses the /openai/v1/files content route.","sources":["s435"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/azure/files.rb"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"Chat Completions batches have a create, poll, results and cancel lifecycle under /openai/v1/batches, always through the Chat Completions protocol. They need a Global or Data Zone Batch deployment. No Azure live recording covers them.","sources":["s15","s16"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/chat_completions/batches.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/chat_completions/batches_spec.rb","notes":"Covers the Azure batch URLs and endpoint name."}]},"embedding_batches":{"v1":"missing","v2":"na","offered":false,"notes":"The batch service rejects other input URLs: a file with /v1/embeddings returned url_mismatch and listed /v1/responses and /v1/chat/completions as the only valid batch URLs on the audited Foundry resource, whatever the batch reference and v1 enum list. RubyLLM has no Azure embedding batch adapter.","sources":["s707","s16","s15","s434"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embedding_batches.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"Cohere-embed-v3-english","notes":"POST /openai/v1/batches rejected an uploaded /v1/embeddings file with url_mismatch. The file was then deleted."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated advisor tool operation is documented in the Azure OpenAI v1 or Foundry REST inventories. The Foundry Claude pages do not list one. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Perplexity Agent API, which is not an API of this provider.","sources":["s149","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"agent_skills":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Azure documents versioned skills mounted on the hosted shell environment. A raw shell tool definition can reference them through with_provider_tools. RubyLLM has no API for Azure skill upload or versions; Chat#with_skills is a separate local feature.","sources":["s708","s709"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"}]},"async_research":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure documents deep research with o3-deep-research in background mode. Raw fields can start a job, but RubyLLM has no retrieval, polling or cancellation on Azure and sends store false by default.","sources":["s710","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"}]},"background_responses":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure documents background Responses with polling and cancellation. Raw fields pass through, but RubyLLM has no retrieval, polling or cancellation lifecycle and sends store false by default.","sources":["s8","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"}]},"browser_use":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure documents browser automation through the computer tool and Playwright, driven by your own code. A raw tool definition can be sent, but RubyLLM does not implement the action and screenshot loop. Limited access applies to gpt-5.4.","sources":["s711","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"}]},"classification":{"v1":"na","v2":"native","offered":true,"notes":"judge uses the Azure Decisions API at /openai/v1/decisions. Choice, predicate and score questions classify text and images into typed answers. gpt-6-luna is the only supported model.","sources":["s712"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb","method":"resolve_protocol"},{"version":"v2","path":"lib/ruby_llm/providers/azure/decisions.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/decisions.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/decisions_spec.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-05","model":"gpt-6-luna","spec":"spec/ruby_llm/providers/azure/decisions_spec.rb","example":"RubyLLM::Providers::Azure::Decisions with the Decisions API judges all three question types through the compact DSL","cassette":"spec/fixtures/vcr_cassettes/providers_azure_decisions_with_the_decisions_api_judges_all_three_question_types_through_the_compact_dsl.yml","notes":"HTTP 200 from /openai/v1/decisions with predicate, choice and score questions."}]},"computer_use":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure documents the computer tool, with limited access for gpt-5.4 and computer-use-preview in the catalog. A raw tool definition can be sent, but RubyLLM does not implement the client action and output loop.","sources":["s711","s696","s704"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"}]},"context_editing":{"v1":"na","v2":"missing","offered":true,"notes":"Claude on Foundry documents context editing through the Messages API beta header. RubyLLM Azure has no Messages route, so raw options cannot reach it.","sources":["s701"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated contextual embeddings operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"dedicated_transcription_api":{"v1":"partial","v2":"native","offered":true,"notes":"The audio/transcriptions route is the same integration as Transcription, including streamed and detailed results. No Azure live recording covers it.","sources":["s149","s312"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/audio.rb"},{"version":"v2","path":"spec/ruby_llm/providers/azure/media_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/media_spec.rb","notes":"Covers resource, deployment and /openai/v1 routes and multipart handling."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated dubbing operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"fast_inference":{"v1":"na","v2":"passthrough","offered":true,"notes":"Azure documents priority processing and Flex processing, selected on eligible deployments or per request with service_tier. with_provider_options can send service_tier; RubyLLM has no dedicated setting.","sources":["s713","s714"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Azure documents file listing and deletion. RubyLLM Files supports upload, metadata and download only.","sources":["s435"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/azure/files.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Azure vector stores are the file search stores, with a full create, index and search lifecycle. RubyLLM can only use a prepared store with file_search.","sources":["s699"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/azure/files.rb"}]},"fill_in_middle":{"v1":"na","v2":"unknown","offered":null,"notes":"The Foundry completions reference documents suffix only for gpt-3.5-turbo-instruct, which the Azure catalogs and retirement schedule do not list, and Codestral is listed without a documented fill-in-the-middle route. Whether any deployment accepts it is unconfirmed. gpt-35-turbo-instruct-0914 appears in GET /openai/v1/models but is not deployed on the audited resource, so no live check ran; a suffix sent to gpt-5-nano returned 400. RubyLLM has no completions route.","sources":["s715","s704","s716","s702"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"gpt-5-nano","notes":"POST /openai/v1/completions with a suffix returned 400: The requested operation is unsupported."}]},"fine_tuning":{"v1":"missing","v2":"missing","offered":true,"notes":"Azure documents fine-tuning jobs. RubyLLM has no job lifecycle; calling a fine-tuned deployment works as chat. Training APIs are a scope boundary.","sources":["s717"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google Maps grounding, which is not an API of this provider.","sources":["s149","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"guardrail_configuration":{"v1":"na","v2":"passthrough","offered":true,"notes":"Azure documents selecting a custom guardrail per request with the x-policy-id header, which with_headers can send. Creating guardrails happens in the portal or the management API, which RubyLLM does not cover.","sources":["s718","s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_headers"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Azure documents stored responses, previous_response_id chaining and a Conversations API. Claude on Foundry documents message threads. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s8","s719","s701"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch service rejects other input URLs: a file with /v1/embeddings returned url_mismatch and listed /v1/responses and /v1/chat/completions as the only valid batch URLs on the audited Foundry resource, whatever the batch reference and v1 enum list. RubyLLM has no image batch adapter.","sources":["s707","s15","s434"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /openai/v1/batches returned url_mismatch and named only Responses and Chat Completions as valid batch URLs. No image file was submitted: the result rests on that allow-list, from an uploaded /v1/embeddings file."}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Interactions API, which is not an API of this provider.","sources":["s149","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"response_format json_object goes through with_provider_options. Structured schemas have their own public API.","sources":["s720"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names Amazon Bedrock Mantle, which is not an API of this provider.","sources":["s149","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"manual_compaction":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#compact calls Responses /compact and keeps the full history locally, resetting only later wire context. Pending tool calls block compaction.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"apply_compaction"},{"version":"v2","path":"lib/ruby_llm/chat.rb"},{"version":"v2","path":"lib/ruby_llm/active_record/chat_methods.rb"},{"version":"v2","path":"lib/ruby_llm/agent.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with azure","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_azure.yml","notes":"HTTP 200 from /openai/v1/responses/compact returning response.compaction, then a continued response."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated music generation operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch service rejects other input URLs: a file with /v1/embeddings returned url_mismatch and listed /v1/responses and /v1/chat/completions as the only valid batch URLs on the audited Foundry resource, whatever the batch reference and v1 enum list. That leaves completions, moderations, images and videos out, despite the enum. RubyLLM has no adapter for them.","sources":["s707","s15"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /openai/v1/batches returned url_mismatch and named only Responses and Chat Completions as valid batch URLs. No completion, moderation, image or video file was submitted: the result rests on that allow-list, from an uploaded /v1/embeddings file."}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Omni video endpoint, which is not an API of this provider.","sources":["s149","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"partner_model_protocols":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure declares a Cohere embedding and rerank protocol, selected by operation, known model names or azure_protocol: :cohere. The Claude Messages API at /anthropic/v1/messages has no Azure route in RubyLLM. Grok and DeepSeek use the OpenAI-compatible routes.","sources":["s701","s721","s152","s396","s722"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/cohere.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/rerank.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"Cohere-embed-v3-english","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","example":"RubyLLM::Providers::Azure::Cohere embeds an image through the active Azure Cohere endpoint with billed usage","cassette":"spec/fixtures/vcr_cassettes/providers_azure_cohere_embeds_an_image_through_the_active_azure_cohere_endpoint_with_billed_usage.yml","notes":"HTTP 200 from /providers/cohere/v2/embed for one image."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/azure/cohere_spec.rb","notes":"Covers URL normalization, authentication, image limits, routing and typed rerank results."}]},"prefix_completion":{"v1":"na","v2":"unknown","offered":null,"notes":"The Foundry reference lists /openai/v1/completions, but no Azure page documents a prefix or assistant prefill behavior. A bare endpoint does not establish it. No Claude model is deployed on the audited resource; a trailing assistant message returned empty content on Kimi-K2.5 and restarted the answer on Grok 4.1 Fast, neither of which settles it. RubyLLM has no completions route.","sources":["s715","s701"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"grok-4-1-fast-non-reasoning","notes":"A final assistant turn on /openai/v1/chat/completions was not continued."}]},"programmatic_tool_calling":{"v1":"na","v2":"partial","offered":true,"notes":"Azure documents programmatic tool calling on the Responses API. Raw tool options and the programmatic_tool_calling tool can be sent, but RubyLLM does not implement caller information or resuming the generated program. Supported deployments are not listed.","sources":["s723"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching translates key and ttl to prompt_cache_key and prompt_cache_options on GPT-5.6 and later. prompt_cache_retention on earlier models goes through with_provider_options.","sources":["s11"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"}]},"provider_server_fallback":{"v1":"na","v2":"passthrough","offered":true,"notes":"Azure documents automatic failover in model router, selected by deployment name, and per-request spillover with the x-ms-spillover-deployment header, which with_headers can send. RubyLLM has no dedicated setting, and with_fallbacks is a separate client feature.","sources":["s724","s725"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_headers"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Azure documents the GPT Realtime API and GPT-Live. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s726"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated response caching operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"Azure::Responses serves /openai/v1/responses with explicit selection and automatic routing for deployment names matching recent GPT ids. Other deployment names need protocol: :responses.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-5.6-luna","spec":"spec/ruby_llm/providers/azure/responses_spec.rb","example":"RubyLLM::Providers::Azure::Responses a GPT-5.6 deployment routes to Responses without being asked","cassette":"spec/fixtures/vcr_cassettes/providers_azure_responses_a_gpt-5_6_deployment_routes_to_responses_without_being_asked.yml","notes":"The deployment name routes to /openai/v1/responses."}]},"responses_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"The batch guide documents /v1/responses input. RubyLLM registers no Responses batch dialect for Azure.","sources":["s15"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/chat_completions/batches.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated standalone search operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"server_apply_patch":{"v1":"missing","v2":"partial","offered":true,"notes":"The Responses REST schema defines the apply_patch tool and its calls and outputs; no Azure guide covers it. A raw tool definition can be sent, but RubyLLM does not implement the client execution loop.","sources":["s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"}]},"server_tool_image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"The image_generation alias works on the Responses route with GPT image deployments. Azure does not support multi-turn editing or streaming here. No Azure live recording covers it.","sources":["s8"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/providers/azure/responses.rb"}]},"server_tool_search":{"v1":"missing","v2":"native","offered":true,"notes":"Deferred tools from with_tools(defer:) send defer_loading with the tool_search tool and namespaces on the Responses route, and searched calls return to their namespace. Azure supports it on gpt-5.4 and later. No Azure live recording covers it.","sources":["s727"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_tool_search"},{"version":"v2","path":"lib/ruby_llm/chat/tool_search.rb"},{"version":"v2","path":"spec/ruby_llm/protocols/responses/tools_tool_search_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/responses/tools_tool_search_spec.rb","notes":"Covers the shared Responses tool search rendering."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated X search tool operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure says live search options do not apply to Grok deployments. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s722","s696"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"shell_computer_tools":{"v1":"missing","v2":"partial","offered":true,"notes":"Azure documents hosted and local shell. A raw hosted shell definition can be sent and its items are kept as server tool calls. Local shell needs a client execution loop RubyLLM does not implement.","sources":["s709","s711"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated sound effects operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak yields SpeechChunk bytes and returns the complete Speech. Azure documents stream_format audio and sse; RubyLLM reads the binary stream. No Azure live recording covers it.","sources":["s149","s342"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb"},{"version":"v2","path":"spec/ruby_llm/speech_streaming_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/speech_streaming_spec.rb","notes":"Covers the Azure deployment speech URL with a streamed block."}]},"task_budgets":{"v1":"na","v2":"missing","offered":true,"notes":"Claude on Foundry lists task_budget token budgets in beta on recent models. RubyLLM Azure has no Messages route, so raw options cannot reach it.","sources":["s701"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated text analysis operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Azure documents vector stores, their files, batches and search. RubyLLM has no API for them.","sources":["s699"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/azure/files.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The batch service rejects other input URLs: a file with /v1/embeddings returned url_mismatch and listed /v1/responses and /v1/chat/completions as the only valid batch URLs on the audited Foundry resource, whatever the batch reference and v1 enum list. RubyLLM has no video batch adapter.","sources":["s707","s15","s434"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/videos.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /openai/v1/batches returned url_mismatch and named only Responses and Chat Completions as valid batch URLs. No video file was submitted: the result rests on that allow-list, from an uploaded /v1/embeddings file."}]},"video_editing":{"v1":"missing","v2":"missing","offered":true,"notes":"The Foundry REST reference documents video remix and edits on /openai/v1/videos. RubyLLM Azure uses only the jobs route, which has no remix or edit operation.","sources":["s728","s729","s14"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/videos.rb"}]},"video_extension":{"v1":"na","v2":"missing","offered":true,"notes":"The Foundry REST reference documents POST /openai/v1/videos/extensions. RubyLLM Azure has no extension route.","sources":["s730","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure/videos.rb"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice cloning operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice conversion operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated standalone web fetch operation is documented in the Azure OpenAI v1 or Foundry REST inventories. Azure Speech, Content Safety, Document Intelligence, AI Search and Foundry Agent Service are separate products.","sources":["s149","s696","s700"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"}]},"websockets":{"v1":"missing","v2":"missing","offered":true,"notes":"Azure documents Responses over WebSocket, continued with previous_response_id and store false. RubyLLM has no WebSocket transport for Responses.","sources":["s153"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/providers/azure.rb"},{"version":"v2","path":"lib/ruby_llm/transport/websocket_connection.rb"}]}},"gaps":[{"notes":"Partner model protocols: RubyLLM Azure covers Cohere embeddings and reranking. The Claude Messages API on Foundry has no Azure route.","sources":["s701","s152"]},{"notes":"Token counting, Context editing, Task budgets: Foundry documents these for Claude through the Messages API, which RubyLLM Azure does not reach.","sources":["s701"]},{"notes":"OCR / document parsing: Foundry also offers Cohere Parse, and RubyLLM Azure has no route for it.","sources":["s704","s152"]},{"notes":"Video editing, Video extension: The Foundry REST reference documents remix, edits and extensions on /openai/v1/videos. RubyLLM Azure uses only the jobs route.","sources":["s728","s729","s730"]},{"notes":"Programmatic tool calling, Browser use, Computer use, Apply patch tool, Shell and computer tools: Raw tool definitions can be sent, but RubyLLM does not implement their client execution and output loops.","sources":["s723","s711","s709","s696"]},{"notes":"Asynchronous research, Background Responses jobs: Raw fields can start a job, but RubyLLM has no retrieval, polling or cancellation lifecycle and sends store false by default.","sources":["s710","s8"]},{"notes":"File listing and deletion, File search store management, Vector store management: Azure documents these lifecycles. RubyLLM Files supports upload, metadata and download only.","sources":["s435","s699"]},{"notes":"Fine-tuning: RubyLLM has no job lifecycle. Training APIs are a scope boundary.","sources":["s717"]},{"notes":"Responses batches: Azure documents /v1/responses batch input, but RubyLLM registers no Responses batch dialect for Azure.","sources":["s15"]},{"notes":"Responses over WebSocket: Azure documents this mode. RubyLLM has no WebSocket transport for Responses.","sources":["s153"]},{"notes":"Agent skills, Fast inference, Guardrail configuration, Provider-managed fallback, JSON object mode: These need raw options or headers, with no dedicated RubyLLM setting.","sources":["s708","s713","s718","s725","s720"]},{"notes":"Fill-in-the-middle completion, Prefix completion: the audited resource has no gpt-35-turbo-instruct or Claude deployment, so these remain unverified.","sources":["s715","s704"]}]},{"id":"xai","name":"xAI","notes":"Covers the public xAI inference, Files and Collections APIs at api.x.ai and management-api.x.ai, documented at docs.x.ai under the SpaceXSI name. Consumer Grok, Grok Build and Grok Bot are separate products.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.chat uses the Responses API by default. protocol: :chat_completions selects Chat Completions, which xAI files under legacy. Multi-agent models work only on Responses.","sources":["s731","s732","s733","s157"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality xai/grok-4-1-fast-non-reasoning can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_xai_grok-4-1-fast-non-reasoning_can_have_a_basic_conversation.yml","notes":"Answers on the Responses endpoint. The spec names a retired slug that the registry aliases to grok-4.3."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"A block on ask streams server-sent events on Responses or Chat Completions.","sources":["s734","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses xai/grok-4-1-fast-non-reasoning supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_xai_grok-4-1-fast-non-reasoning_supports_streaming_responses.yml","notes":"Server-sent events on the Responses endpoint."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends text.format with json_schema on Responses and response_format on Chat Completions.","sources":["s736","s735","s732"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"schema_format"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat #with_schema with xai/grok-4-1-fast-non-reasoning accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_xai_grok-4-1-fast-non-reasoning_accepts_a_json_schema_and_returns_structured_output.yml","notes":"The request carries text.format with type json_schema."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments go as input_image on Responses and image_url on Chat Completions. xAI lists JPEG and PNG, up to 20 MiB each.","sources":["s737","s18"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_image"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat.rb","method":"supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models xai/grok-4-1-fast-non-reasoning can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_xai_grok-4-1-fast-non-reasoning_can_understand_local_images.yml","notes":"A local image goes as an input_image data URI."}]},"pdf_input":{"v1":"missing","v2":"native","offered":true,"notes":"On Responses, a PDF URL goes as input_file with file_url, an uploaded file with file_id, and a local PDF inline with file_data. Attached files turn the request into an agentic document search. The Chat Completions dialect rejects document attachments. Only the file_url path has a recording.","sources":["s738","s19","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_document"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_provider_file"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"document_urls?"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat.rb","method":"document_attachments"}],"verification":[{"status":"passed","recorded_on":"2026-10-25","model":"grok-4.3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat document URL models xai/grok-4-1-fast-non-reasoning reads PDFs the provider fetches from their URL","cassette":"spec/fixtures/vcr_cassettes/chat_document_url_models_xai_grok-4-1-fast-non-reasoning_reads_pdfs_the_provider_fetches_from_their_url.yml","notes":"The request carries input_file with file_url. RubyLLM does not download the PDF."}]},"audio_input":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"xAI takes conversational audio through its Speech to Speech realtime API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application. The Responses and Chat Completions references document text, image and file parts only. An InputAudio schema in the OpenAPI file is not referenced by any request.","sources":["s366","s20","s367","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"supported_attachment?"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The Responses and Chat Completions inputs document text, image and file parts, not video. X Search can inspect videos it finds through a tool option; that is a server-search capability.","sources":["s735","s739","s740"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Reasoning summaries and encrypted reasoning from Responses become Message#thinking and replay on later turns. RubyLLM asks for reasoning.encrypted_content on every request. grok-4.7 always returns it.","sources":["s741","s18"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_reasoning_summary"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_reasoning_signature"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_reasoning_item"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking xai/grok-3-mini returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_xai_grok-3-mini_returns_thinking_when_available.yml","notes":"The response carries a reasoning summary. The spec names a retired slug that the registry aliases to grok-4.3."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning.effort. xAI documents low, medium and high on grok-4.5, grok-4.6 and grok-4.7, and xhigh from grok-4.6. Reasoning cannot be disabled on those models.","sources":["s741","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"URL annotations and collections:// source references become citations, including those that arrive only in the completed stream event.","sources":["s742","s743"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"spec/ruby_llm/providers/xai/responses_spec.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with xai/grok-4.3 searches, cites, and counts the sources it used","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_xai_grok-4_3_searches_cites_and_counts_the_sources_it_used.yml","notes":"The response carries url_citation annotations that become citations."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools work on Responses and Chat Completions. Multi-agent models reject client-side function tools.","sources":["s744","s733"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling xai/grok-4-1-fast-non-reasoning can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_xai_grok-4-1-fast-non-reasoning_can_use_tools.yml","notes":"A function_call item runs locally and its result goes back in the next request."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) sends tool_choice, including a specific function.","sources":["s744","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control xai/grok-4-1-fast-non-reasoning respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_xai_grok-4-1-fast-non-reasoning_respects_specific_tool_choice.yml","notes":"The model calls the tool the request names."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"One response can carry several function calls, and RubyLLM runs them all.","sources":["s744"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_tools"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling xai/grok-4-1-fast-non-reasoning can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_xai_grok-4-1-fast-non-reasoning_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"One response carries several function_call items."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) sends parallel_tool_calls false.","sources":["s744","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parallel_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control xai/grok-4-1-fast-non-reasoning respects calls one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_xai_grok-4-1-fast-non-reasoning_respects_calls_one_for_sequential_execution.yml","notes":"The response echoes parallel_tool_calls false."}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends the web_search tool. Domain filters and image options pass as alias options.","sources":["s745","s17"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with xai/grok-4.3 searches, cites, and counts the sources it used","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_xai_grok-4_3_searches_cites_and_counts_the_sources_it_used.yml","notes":"Search calls, cited pages and web_search_requests usage are all present."}]},"server_web_fetch":{"v1":"passthrough","v2":"native","offered":true,"notes":"xAI has no separate fetch tool: browsing pages is part of web_search, which RubyLLM registers. There is no web_fetch alias.","sources":["s745"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:code_interpreter) sends the code_interpreter type xAI documents for Responses. The :code_execution alias sends type code_execution, which is the xAI SDK name and is not in the documented Responses tool types. No recording covers either alias.","sources":["s746","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}]},"server_file_search":{"v1":"missing","v2":"native","offered":true,"notes":"The file_search and collections_search aliases send file_search with options such as the collection ids. RubyLLM cannot create the collections; prepare them through xAI.","sources":["s743","s747"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"The mcp alias takes a portable name and url plus options such as allowed_tools. xAI runs the remote tools and returns server tool calls. xAI does not support require_approval or connector_id.","sources":["s339","s17"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP execution xai executes a read-only tool and replays its history","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_execution_xai_executes_a_read-only_tool_and_replays_its_history.yml","notes":"xAI runs a read-only MCP tool, and the follow-up request replays the typed results."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"xAI caches matching prefixes automatically, and cached counts land in cache_read. with_caching(key:) sends prompt_cache_key, which xAI treats as x-grok-conv-id. The header itself needs with_headers.","sources":["s21","s748","s732","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"apply_prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_usage"}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"xAI documents automatic prefix caching and routing keys, with no caller-selected boundaries. The deprecated Messages endpoint marks cache_control as unsupported.","sources":["s21","s749","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"inject_cache_breakpoint"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"xAI documents automatic prefix caching and routing keys, not separately created cache resources.","sources":["s21","s748"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"compaction":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#compact calls POST /v1/responses/compact and continues from the returned compaction item. It needs the Responses protocol. with_compaction has no xAI wire option: the reference says context_management is parsed but not yet executed, so RubyLLM drops it.","sources":["s22","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"COMPACTION_PROVIDERS"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"compact"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"Compaction"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with xai","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_xai.yml","notes":"Chat#compact calls responses/compact, and the next request continues from the compaction item."}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.tokenize calls POST /v1/tokenize-text for plain text. xAI documents no endpoint that counts a full request, so Chat#count_tokens is not available. Token ids are not billed usage.","sources":["s23"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/tokenization.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"Tokenization"},{"version":"v2","path":"lib/ruby_llm/tokenization.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/tokenization_spec.rb","example":"RubyLLM::Tokenization tokenizes text with xai through the public API","cassette":"spec/fixtures/vcr_cassettes/tokenization_tokenizes_text_with_xai_through_the_public_api.yml","notes":"Token ids come back from tokenize-text and the count matches them."}]},"image_generation":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.paint calls images/generations and parses several images. xAI documents aspect_ratio, resolution and quality, which pass through provider_options. The size keyword is sent as size, a field the Images reference does not list.","sources":["s750","s751"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb","method":"render_image_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/images.rb","method":"render_image_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/images.rb","method":"parse_image_responses"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-imagine-image","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality xai/grok-imagine-image can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_xai_grok-imagine-image_can_paint_images.yml","notes":"The request omits size, and the response returns an image URL."}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"with: selects images/edits and sends JSON references for URLs and local images. xAI takes up to five images and documents no mask, so RubyLLM rejects one. Uploaded files are sent in the url field; xAI documents them under file_id.","sources":["s752","s753","s751","s345"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb","method":"render_edit_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/images.rb","method":"render_edit_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/images.rb","method":"image_reference_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-03","model":"grok-imagine-image-2.0","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image edit functionality with xAI reference images xai/grok-imagine-image-2.0 supports image edits with reference images","cassette":"spec/fixtures/vcr_cassettes/image_edit_functionality_with_xai_reference_images_xai_grok-imagine-image-2_0_supports_image_edits_with_reference_images.yml","notes":"A local image goes as a data URI in the images array on images/edits."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate does text-to-video and image-to-video with job polling and Video results. Reference-to-video fields such as reference_images and reference_audios, and keyframes, pass through provider_options; with: takes one reference.","sources":["s754","s755"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"render_video_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"parse_video_job_status"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-imagine-video","spec":"spec/ruby_llm/video_spec.rb","example":"RubyLLM::Video basic functionality xai/grok-imagine-video can animate videos","cassette":"spec/fixtures/vcr_cassettes/video_basic_functionality_xai_grok-imagine-video_can_animate_videos.yml","notes":"The job is created, polled until done, and the MP4 is downloaded."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak calls POST /v1/tts with voice_id, language auto and output_format. It returns a Speech with save and to_blob.","sources":["s347","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/speech.rb","method":"render_speech_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-tts","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality xai/grok-tts can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_xai_grok-tts_can_speak.yml","notes":"POST /v1/tts returns audio bytes."}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe calls POST /v1/stt with a multipart file. grok-stt maps to the API name grok-voice-transcribe-2.0. Options such as format and keyterm pass through provider_options.","sources":["s160","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/transcription.rb","method":"render_transcription_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-02","model":"grok-stt","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality xai/grok-stt can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_xai_grok-stt_can_transcribe_audio.yml","notes":"The multipart request names grok-voice-transcribe-2.0 and returns the transcript."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe opens the stt WebSocket and streams WAV PCM or G.711 audio, mono or multichannel. Raw Opus and other containers are not converted.","sources":["s20","s160","s358"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/streaming_transcription.rb","method":"stream_transcription"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"StreamingTranscription"}],"verification":[{"status":"passed","recorded_on":"2026-11-02","model":"grok-stt","spec":"spec/ruby_llm/protocols/xai/streaming_transcription_spec.rb","example":"RubyLLM::Protocols::XAI::StreamingTranscription streams transcription through the public API with typed chunks and word timing","cassette":"spec/fixtures/websocket_cassettes/transcription_xai.json","notes":"Partial events, word timing and a final transcript come over the stt WebSocket."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"speaker_names: on transcribe sends diarize, and each word carries a numeric speaker index. xAI has no named speakers, so speaker_references and named_speakers are rejected.","sources":["s160","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/transcription.rb","method":"rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-stt","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality xai/grok-stt labels words with speakers when speaker names are given","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_xai_grok-stt_labels_words_with_speakers_when_speaker_names_are_given.yml","notes":"The request sets diarize, and each word carries a speaker index."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no moderation endpoint. The guidelines request field changes how Grok writes its answer and returns no verdict; see Guardrail configuration.","sources":["s756","s739","s757"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no OCR or document parsing endpoint. Chat with Files searches documents through a chat request, which is not an OCR API.","sources":["s756","s739","s738"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"embeddings":{"v1":"na","v2":"unknown","offered":null,"notes":"The OpenAPI file lists POST /v1/embeddings and GET /v1/embedding-models. GET /v1/embedding-models answers with an empty list for this key, and POST /v1/embeddings returned the same does not exist or no access 404 for grok-embedding-small as for a made-up name, so the routes exist and whether any team gets a model is unconfirmed. The REST reference, model catalog and pricing page list no embedding model, and a Collections example names grok-embedding-small for indexing only. RubyLLM.embed reaches the inherited endpoint.","sources":["s739","s756","s18","s758"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"grok-embedding-small","notes":"GET /v1/embedding-models returned 200 with an empty list. POST /v1/embeddings returned 404: the model does not exist or the team has no access."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The documented embedding input takes strings and token arrays only, not images.","sources":["s739","s756"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no reranking endpoint.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload calls POST /v1/files with purpose and expires_in. xAI needs its fields before the file part, and RubyLLM sends them in that order. The uri option is rejected.","sources":["s19","s759"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/files.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads xai uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_xai_uploads_a_pdf_through_the_files_api.yml","notes":"POST /v1/files returns the file id and byte size."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download returns a DownloadedFile with save and to_blob, and UploadedFile.find reads metadata. No recording covers xAI downloads.","sources":["s759","s760"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"download"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find"},{"version":"v2","path":"lib/ruby_llm/downloaded_file.rb"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch runs the xAI lifecycle: create, add requests, poll, collect results, cancel. Not every model accepts batches. xAI batches can take a long time: on November 10, 2026 they stayed pending for over 30 minutes.","sources":["s26","s761","s18"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb","method":"create_batch"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb","method":"batch_results"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb","method":"batch_protocol"}],"verification":[{"status":"unavailable","checked_on":"2026-11-10","model":"grok-4-1-fast-non-reasoning","notes":"Two batches stayed pending for over 30 minutes with no request processed, so the live recording from October 11, 2026, which passed, was removed until xAI batches finish again."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Batch API accepts chat, Responses, image and video requests, and does not list embeddings.","sources":["s26","s761"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated advisor tool operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Perplexity Agent API; it is not an API of this provider.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The OpenAPI file lists /v1/skills, but GET /v1/skills returned 404, the same as a nonsense path, and the guides and REST reference do not describe hosted skills. RubyLLM has no integration.","sources":["s739","s756","s157"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/skills returned 404 Not Found, as did a nonsense path."}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"xAI documents multi-agent research as a synchronous or streamed Responses model. Deferred completions and batches are general queues, and background is marked unsupported, so no asynchronous research operation is documented.","sources":["s733","s735","s159","s756"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Responses reference labels background unsupported. Deferred Chat Completions is a different API and does not establish Responses background support.","sources":["s735","s159"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated browser use operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated classification operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no computer use tool. Grok Bot computers are a separate product. The separate local shell tool type is covered under Shell and computer tools.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated context editing operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"contextual_embeddings":{"v1":"na","v2":"missing","offered":true,"notes":"Collections can prepend metadata fields to each embedding chunk with inject_into_chunk, through the Management API. RubyLLM does not manage collections.","sources":["s762","s747","s162"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe uses the dedicated stt endpoint. This row overlaps Transcription.","sources":["s160","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/transcription.rb","method":"transcription_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-02","model":"grok-stt","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality xai/grok-stt can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_xai_grok-stt_can_transcribe_audio.yml","notes":"The multipart request names grok-voice-transcribe-2.0 and returns the transcript."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated dubbing operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Priority processing uses service_tier on Chat Completions and Responses. with_provider_options(service_tier: \"priority\") forwards it. There is no dedicated priority API, and the served tier stays in the raw response.","sources":["s161","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"provider_options"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"xAI documents file listing, deletion, renaming and public URLs. RubyLLM has upload, metadata and download only.","sources":["s759","s760"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/files.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"xAI Collections are the file search stores, with creation, document and deletion APIs behind a Management API key. RubyLLM can query an existing collection through the hosted tool, and cannot manage one.","sources":["s162","s758"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/files.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The legacy completions endpoint marks suffix as unsupported, and xAI documents no other fill-in-the-middle operation.","sources":["s749","s739","s756"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The current inference and management inventory documents no fine-tuning job endpoint. This means no documented hosted offering, not that private enterprise arrangements are impossible.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Maps grounding; it is not an API of this provider.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"guardrail_configuration":{"v1":"na","v2":"passthrough","offered":true,"notes":"A guidelines array on Responses or Chat Completions sets per-request behavior for the nsfw, violence and profanity categories. with_provider_options(guidelines: [...]) forwards it. RubyLLM has no dedicated option. xAI rejects guidelines on multi-agent models and deferred completions.","sources":["s757","s27","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"provider_options"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"xAI stores responses for 30 days and continues them through previous_response_id, with read and delete endpoints. This is outside RubyLLM 2.0 coverage; conversation state stays in the application, and RubyLLM sends store false.","sources":["s731","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"image_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"The Batch API takes image generation and editing requests. RubyLLM.batch takes chats and embedding requests only.","sources":["s26","s27"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/batch.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Interactions API; it is not an API of this provider.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"JSON object mode needs a raw option: text.format on Responses or response_format on Chat Completions, through with_provider_options. with_schema is the public method for schemas. The xAI guides describe schemas, and the OpenAPI file lists json_object.","sources":["s739","s736"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_request_options_spec.rb","example":"RubyLLM::Chat with params xai/grok-4-1-fast-non-reasoning supports response_format param","cassette":"spec/fixtures/vcr_cassettes/chat_with_params_xai_grok-4-1-fast-non-reasoning_supports_response_format_param.yml","notes":"Raw text.format with type json_object goes through with_provider_options and returns JSON."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Amazon Bedrock Mantle; it is not an API of this provider.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"manual_compaction":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#compact calls POST /v1/responses/compact. The returned item replaces the wire context for later requests, while the local transcript keeps every message. Pending tool calls block compaction.","sources":["s22","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"compact"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"Compaction"},{"version":"v2","path":"lib/ruby_llm/active_record/chat_methods.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with xai","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_xai.yml","notes":"Chat#compact calls responses/compact, and the next request continues from the compaction item."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated music generation operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"nonchat_batches":{"v1":"na","v2":"missing","offered":true,"notes":"The Batch API also takes image generation, image editing, video generation, video editing and video extension requests. RubyLLM.batch takes chats and embedding requests only. This row overlaps Image batches and Video batches.","sources":["s26"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/batch.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Omni video endpoint; it is not an API of this provider.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated partner model protocols operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"Chat Completions and Responses do not continue a final assistant message: requests ending in an assistant prefill were answered from the start on grok-4.3, and on grok-4.20 through Chat Completions. The legacy completions endpoint continues a raw prompt on non-reasoning models, with no prefix parameter. RubyLLM has no completions operation.","sources":["s749","s732","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"grok-4.3","notes":"A trailing assistant \"Zaffre, then\" was ignored in two Chat Completions requests and one Responses request; the reply listed three unrelated colours."}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated programmatic tool calling operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching(key:) sends prompt_cache_key on Responses and Chat Completions. xAI uses it for routing, and it does not set cache boundaries. The x-grok-conv-id header itself needs with_headers.","sources":["s748","s735","s732"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"apply_prompt_cache_params"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated provider-managed fallback operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"xAI documents realtime voice conversations over WebSocket. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s20","s366","s367"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated response caching operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"The xAI Responses dialect is the default protocol. RubyLLM sends store false and replays local history.","sources":["s735","s731","s17"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-4.3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat responses protocol dialects with xai/grok-4.3 chats on the Responses endpoint by default","cassette":"spec/fixtures/vcr_cassettes/chat_responses_protocol_dialects_with_xai_grok-4_3_chats_on_the_responses_endpoint_by_default.yml","notes":"Two turns on the responses path, replaying local history with store false."}]},"responses_batches":{"v1":"missing","v2":"native","offered":true,"notes":"Chats on Responses go into a batch under the responses key. Results come back as chat completions, which RubyLLM parses in submission order. xAI batches can take a long time: on November 10, 2026 they stayed pending for over 30 minutes.","sources":["s26","s761"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb","method":"batch_request_type"},{"version":"v2","path":"lib/ruby_llm/providers/xai/chat_completions/batches.rb","method":"parse_batch_result"}],"verification":[{"status":"unavailable","checked_on":"2026-11-10","model":"grok-4-1-fast-non-reasoning","notes":"Two batches stayed pending for over 30 minutes with no request processed, so the live recording from October 11, 2026, which passed, was removed until xAI batches finish again."}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no standalone web search API. The one search endpoint, POST /v1/documents/search, searches your collections and is covered under Vector store management.","sources":["s756","s763","s157"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated apply patch tool operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"server_tool_image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:image_generation) sends the image_generation tool. No recording covers it.","sources":["s764","s17"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"}]},"server_tool_search":{"v1":"na","v2":"unknown","offered":null,"notes":"Limited to alpha users: a request with tool_search and defer_loading returned 403 on grok-4.3 and grok-4-1-fast-non-reasoning, while the request validator lists tool_search as a tool type. The guides and tools overview do not mention it, and no alpha access was available to test it. RubyLLM gives xAI chats its own client-side tool search instead, so with_tools(defer: true) works on any key.","sources":["s739","s17","s756"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/chat/tool_search.rb","method":"searching_tools_here?"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"grok-4.3","notes":"POST /responses returned 403: The tool_search tool and defer_loading are only available for alpha users. grok-4-1-fast-non-reasoning returned the same."},{"status":"passed","recorded_on":"2026-11-09","model":"grok-4-1-fast-non-reasoning","spec":"spec/ruby_llm/chat_client_tool_search_spec.rb","example":"RubyLLM::Chat on xAI, whose tool search is limited to alpha accounts loads a deferred tool through its own search and calls it","cassette":"spec/fixtures/vcr_cassettes/chat_on_xai_whose_tool_search_is_limited_to_alpha_accounts_loads_a_deferred_tool_through_its_own_search_and_calls_it.yml","notes":"RubyLLM's tool_search function loaded weather_lookup on the same key, and the model called it. No defer_loading in the requests."}]},"server_x_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:x_search) sends the x_search tool, with handle, date and media options as alias options. Usage counts x_search_requests. No recording covers it.","sources":["s740","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/providers/xai/responses.rb","method":"parse_server_tool_use"}]},"shell_computer_tools":{"v1":"na","v2":"missing","offered":true,"notes":"Responses accepts a shell tool with environment { type: \"local\" }, and the model emits shell_call items for your application to run and report back as shell_call_output. The tool appears in the OpenAPI file and not in the guides. No computer tool is documented. RubyLLM has no integration.","sources":["s739","s17","s756"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"grok-4.3","notes":"With tools: [{ type: \"shell\", environment: { type: \"local\" } }], a request to run echo hello returned a shell_call item with commands [\"echo hello\"]."}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated sound effects operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak streams bytes from POST /v1/tts and returns the complete Speech. Timestamp JSON mode is rejected when streaming. The bidirectional wss://api.x.ai/v1/tts endpoint, which streams text in, is not integrated.","sources":["s347","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/speech.rb","method":"stream_speech"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-tts","spec":"spec/ruby_llm/providers/xai/speech_spec.rb","example":"RubyLLM::Providers::XAI::Speech streams speech and retains the complete audio through the public API","cassette":"spec/fixtures/vcr_cassettes/providers_xai_speech_streams_speech_and_retains_the_complete_audio_through_the_public_api.yml","notes":"Typed chunks join into the same bytes as the complete Speech."}]},"task_budgets":{"v1":"na","v2":"passthrough","offered":true,"notes":"max_turns limits the assistant and tool-call turns of one agentic request. with_provider_options(max_turns: 3) forwards it. It limits turns, not tokens, and RubyLLM has no dedicated option.","sources":["s765","s735"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"provider_options"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated text analysis API operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Collections provide the managed vector index, with creation, indexing and POST /v1/documents/search. RubyLLM can query an existing collection through the hosted tool, and cannot manage or search one directly.","sources":["s747","s162","s763"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/files.rb"}]},"video_batches":{"v1":"na","v2":"missing","offered":true,"notes":"The Batch API takes video generation, editing and extension requests. RubyLLM.batch takes chats and embedding requests only.","sources":["s26","s27"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/batch.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb"}]},"video_editing":{"v1":"missing","v2":"native","offered":true,"notes":"A video attachment on with: selects videos/edits. Local files, URLs and uploaded files work, using url or file_id.","sources":["s344","s345","s27"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"video_request_url"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"video_reference"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-imagine-video","spec":"spec/ruby_llm/providers/xai/videos_spec.rb","example":"RubyLLM::Providers::XAI::Videos video editing and extension completes a video edit through the public API","cassette":"spec/fixtures/vcr_cassettes/providers_xai_videos_completes_a_video_edit_through_the_public_api.yml","notes":"The job completes and the MP4 downloads."}]},"video_extension":{"v1":"missing","v2":"native","offered":true,"notes":"extend: selects videos/extensions and takes a Video, URL, path or uploaded file. Duration is the added part, through provider_options. with: and extend: are mutually exclusive.","sources":["s346","s345"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"render_video_extension_payload"},{"version":"v2","path":"lib/ruby_llm/providers/xai/videos.rb","method":"video_extension_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"grok-imagine-video","spec":"spec/ruby_llm/providers/xai/videos_spec.rb","example":"RubyLLM::Providers::XAI::Videos video editing and extension completes a video extension through the public API","cassette":"spec/fixtures/vcr_cassettes/providers_xai_videos_completes_a_video_extension_through_the_public_api.yml","notes":"The job completes and the MP4 downloads."}]},"voice_cloning":{"v1":"missing","v2":"missing","offered":true,"notes":"xAI creates custom voices from a recording in the console, or through POST /v1/custom-voices on Enterprise plans, and only in the United States outside Illinois. RubyLLM has no create, list or delete API. speak accepts a voice id prepared elsewhere.","sources":["s163","s20"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai/speech.rb","method":"render_speech_payload"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice conversion operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated standalone web fetch API operation is documented in the audited API inventory. This is a documentation-scope conclusion, not a claim that the provider or its other products cannot perform related tasks. Scope: The public xAI inference, Files and Collections APIs; excludes Grok consumer app, Grok Build CLI and management APIs. Enterprise-gated inference features remain included with their access restrictions.","sources":["s756","s157","s739"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/providers/xai.rb"}]},"websockets":{"v1":"missing","v2":"missing","offered":true,"notes":"xAI documents Responses over WebSocket on /v1/responses. RubyLLM uses WebSockets for streaming transcription only, and has no Responses transport over them.","sources":["s368","s27"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/xai.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/xai/streaming_transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb"}]}},"gaps":[{"notes":"Contextual embeddings: Collections can inject metadata into embedding chunks through the Management API. RubyLLM does not manage collections.","sources":["s762"]},{"notes":"File listing and deletion, File search store management, Vector store management: xAI exposes file and Collections management and indexing APIs. RubyLLM implements upload, metadata and download, and querying an existing collection through a hosted tool, not collection creation or indexing or file listing or deletion.","sources":["s759","s162"]},{"notes":"Image batches, Video batches, Other batch operations: The Batch API takes image and video requests. RubyLLM.batch takes chats and embedding requests only.","sources":["s26"]},{"notes":"Voice cloning: xAI creates custom voices from a recording, in the console or through an Enterprise-gated API, in the United States outside Illinois. RubyLLM has no custom voice API; speak can use a voice id prepared elsewhere.","sources":["s163"]},{"notes":"Responses over WebSocket: xAI documents Responses over a long-lived WebSocket. RubyLLM has no Responses WebSocket transport.","sources":["s368"]},{"notes":"Fast inference, Guardrail configuration, JSON object mode, Task budgets: service_tier, guidelines, text.format with json_object, and max_turns work only as raw request options.","sources":["s161","s757","s739","s765"]},{"notes":"Code execution tool: The :code_execution alias sends type code_execution, which is not in the documented Responses tool types. Use :code_interpreter, which sends the documented type.","sources":["s746","s739"]},{"notes":"Image editing: Uploaded files are sent in the url field, while xAI documents file_id for them. URLs and local images work.","sources":["s751","s345"]},{"notes":"Video generation: Reference-to-video fields and keyframes need provider_options; with: takes one reference.","sources":["s755"]},{"notes":"Streaming speech generation: The bidirectional text-to-speech WebSocket and timestamps while streaming are not integrated.","sources":["s347"]},{"notes":"Shell and computer tools: the shell tool with a local environment returns shell_call items, and RubyLLM has no integration for them.","sources":["s739","s17","s756"]},{"notes":"Text embeddings, Tool search: the OpenAPI file lists both. No embedding model is available to the audited key, and tool search returned 403 for keys outside the alpha, so availability is unverified. RubyLLM's client-side tool search covers deferred tools on xAI.","sources":["s739","s756"]}]},{"id":"deepseek","name":"DeepSeek","notes":"Scope: the public hosted DeepSeek API at api.deepseek.com. RubyLLM speaks Chat Completions by default, Responses through protocol: :responses, and the Files API. DeepSeek also documents an Anthropic-compatible format at /anthropic, which RubyLLM does not use. Local model hosting, open-weight releases, consumer applications, DeepSeek Harness and third-party deployments are outside scope.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"Chat Completions is the default protocol, and protocol: :responses selects the Responses API. Both serve deepseek-flash and deepseek-v4-pro.","sources":["s766","s166","s165","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality deepseek/deepseek-v4-flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_deepseek_deepseek-v4-flash_can_have_a_basic_conversation.yml","notes":"Recorded against /chat/completions; asserts the answer, role and token usage. DeepSeek reports deepseek-flash as the serving model."},{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat responses protocol dialects with deepseek/deepseek-v4-flash chats with reasoning text on the opt-in Responses protocol","cassette":"spec/fixtures/vcr_cassettes/chat_responses_protocol_dialects_with_deepseek_deepseek-v4-flash_chats_with_reasoning_text_on_the_opt-in_responses_protocol.yml","notes":"Recorded against /responses over two turns."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Both protocols stream. Chat Completions uses chunk deltas and Responses uses semantic events, including reasoning_text deltas.","sources":["s766","s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses deepseek/deepseek-v4-flash supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_deepseek_deepseek-v4-flash_supports_streaming_responses.yml","notes":"Streamed Chat Completions request."},{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat responses protocol dialects with deepseek/deepseek-v4-flash streams reasoning deltas","cassette":"spec/fixtures/vcr_cassettes/chat_responses_protocol_dialects_with_deepseek_deepseek-v4-flash_streams_reasoning_deltas.yml","notes":"Streamed Responses request with reasoning chunks."}]},"structured_output":{"v1":"passthrough","v2":"native","offered":true,"notes":"Select protocol: :responses to use with_schema, which sends a json_schema text format. The default Chat Completions protocol accepts only text and json_object formats, so with_schema raises ArgumentError there. No API recording covers a schema request.","sources":["s166","s165","s766"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"schema_format"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses sends JSON Schema through the Responses text format","notes":"Checks the rendered text.format."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Both protocols send inline and URL images to deepseek-flash; deepseek-v4-pro has no vision. Uploaded file references work only with protocol: :responses, because the Chat Completions dialect raises UnsupportedAttachmentError for them although DeepSeek documents a file content block there.","sources":["s768","s165","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"supported_attachment?"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"render_provider_file"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models deepseek/deepseek-flash can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_deepseek_deepseek-flash_can_understand_local_images.yml","notes":"Sends an image_url part on Chat Completions."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses renders uploaded images as input_image references","notes":"Checks the input_image file_id part on Responses."}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"DeepSeek documents no document input: Responses says file inputs are not supported, and the Anthropic format marks document blocks not supported. The Files API stores images only.","sources":["s166","s769","s168"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"document_attachments"}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The models reference lists only text and image input modalities, and the chat references define no audio content part.","sources":["s770","s766","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The models reference lists only text and image input modalities, and the chat references define no video content part.","sources":["s770","s766","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM parses reasoning_content on Chat Completions and reasoning_text parts on Responses, in full responses and streams, and replays them on later turns. Reasoning token counts reach the thinking bucket.","sources":["s771","s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"parse_reasoning_summary, render_reasoning_item"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"render_thinking"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat DeepSeek thinking control returns reasoning content for effort high","cassette":"spec/fixtures/vcr_cassettes/chat_deepseek_thinking_control_returns_reasoning_content_for_effort_high.yml","notes":"Asserts thinking text on Chat Completions."},{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses continues a streamed reasoning conversation after a tool result","cassette":"spec/fixtures/vcr_cassettes/providers_deepseek_responses_continues_a_streamed_reasoning_conversation_after_a_tool_result.yml","notes":"Asserts streamed reasoning and its replay with the tool result on Responses."}]},"thinking_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_thinking translates effort on both protocols, and effort: :none turns thinking off. On Chat Completions it also sends the thinking toggle. DeepSeek has no token budget, so budget: raises ArgumentError. On Responses, with_thinking(false) sends no reasoning field, so the service default of thinking on applies; use effort: :none.","sources":["s771","s165","s766"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"configure_thinking_payload"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat DeepSeek thinking control disables thinking for effort none","cassette":"spec/fixtures/vcr_cassettes/chat_deepseek_thinking_control_disables_thinking_for_effort_none.yml","notes":"Sends thinking disabled and asserts no reasoning in the response."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"Responses web search on deepseek-v4-pro returned output_text with annotations: [] in two runs. In the run with the full text logged, the sources were written into the answer as Markdown links. No page documents citation output. The Anthropic format ignores citations on input text blocks.","sources":["s166","s165","s769"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_citations"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"deepseek-v4-pro","notes":"A web search through POST /responses returned annotations: [] and the URLs as inline Markdown in the text."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools work on both protocols. In thinking mode RubyLLM replays reasoning with the tool history, as DeepSeek requires. DeepSeek also documents a beta strict mode on its /beta base URL; RubyLLM has no named option for it.","sources":["s772","s771","s166","s766"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"render_thinking"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling deepseek/deepseek-v4-flash can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_deepseek_deepseek-v4-flash_can_use_tools.yml","notes":"Runs the tool loop on Chat Completions."},{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses continues a streamed reasoning conversation after a tool result","cassette":"spec/fixtures/vcr_cassettes/providers_deepseek_responses_continues_a_streamed_reasoning_conversation_after_a_tool_result.yml","notes":"Runs the tool loop on Responses."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) sends none, auto, required or a named tool on both protocols. On Chat Completions, required and named choices return a 400 in thinking mode, so turn thinking off for them.","sources":["s766","s166","s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/capabilities.rb","method":"tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control deepseek/deepseek-v4-flash respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_deepseek_deepseek-v4-flash_respects_specific_tool_choice.yml","notes":"Sends a named choice with thinking disabled."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM runs several tool calls from one response on both protocols. The Responses API always enables parallel calls.","sources":["s165","s766"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_pending_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling deepseek/deepseek-v4-flash can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_deepseek_deepseek-v4-flash_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"The first recorded response carries three tool calls."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"DeepSeek offers no working control. Responses ignores parallel_tool_calls, the Anthropic format ignores disable_parallel_tool_use, and the Chat Completions reference has no such parameter.","sources":["s165","s769","s766"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"server_web_search":{"v1":"na","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) with protocol: :responses sends the web_search tool, which deepseek-v4-pro runs on the server. deepseek-flash ignores the tool. Default Chat Completions has no such tool, and the alias raises UnsupportedServerToolError there. web_search_call items are kept and replayed, and usage.server_tool_use becomes web_search_requests. DeepSeek returns no sources or citations.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"SERVER_TOOL_ALIASES, parse_server_tool_use"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"deepseek-v4-pro","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with deepseek/deepseek-v4-pro through Responses searches, opens the page, counts both, and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_deepseek_deepseek-v4-pro_through_responses_searches_opens_the_page_counts_both_and_continues_from_the_local_transcript.yml","notes":"web_search_call items, counted searches, and a second turn that replays them."},{"status":"passed","recorded_on":"2026-11-10","model":"deepseek-v4-pro","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with deepseek/deepseek-v4-pro through Responses streams a search with its tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_deepseek_deepseek-v4-pro_through_responses_streams_a_search_with_its_tool_calls.yml","notes":"Streamed text matches the final message, with web_search_call items and counts."}]},"server_web_fetch":{"v1":"na","v2":"native","offered":true,"notes":"DeepSeek has no separate fetch tool. with_provider_tools(:web_search) on deepseek-v4-pro with protocol: :responses also opens pages and finds text in them, as open_page actions on ServerToolCall, and usage counts them as web_fetch_requests. There is no :web_fetch alias, as on OpenAI and Azure.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"SERVER_TOOL_ALIASES, parse_server_tool_use"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"deepseek-v4-pro","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with deepseek/deepseek-v4-pro through Responses searches, opens the page, counts both, and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_deepseek_deepseek-v4-pro_through_responses_searches_opens_the_page_counts_both_and_continues_from_the_local_transcript.yml","notes":"An open_page action on the named address, the answer from that page, and web_fetch_requests of 1."}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API lists code_interpreter among built-in tools it ignores, and the reference allows only function, custom and web_search tool types.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API lists file_search among built-in tools it ignores, and the reference allows only function, custom and web_search tool types.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API ignores mcp tools, and the Anthropic format ignores mcp_servers and does not support MCP tool blocks.","sources":["s165","s769"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Context caching is automatic with no request option. RubyLLM reads cache hits from prompt_cache_hit_tokens or cached_tokens on Chat Completions and from cached_tokens on Responses, and bills them as cache reads.","sources":["s773","s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_usage"}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Caching works on prefixes the service detects. The Anthropic format ignores cache_control, and the other formats document no marker.","sources":["s773","s769"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no cache resource endpoint. Caching is automatic, and cached prefixes clear on their own.","sources":["s773","s164"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"Responses does not support context_management or truncation, and a request over the context window returns a 400. No other route documents a provider control that reduces context.","sources":["s165","s164"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"DeepSeek documents a downloadable local tokenizer and an image token calculator, and no hosted counting endpoint in its inventory. RubyLLM raises for count_tokens. Post-request usage is separate.","sources":["s167","s164"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/token_counting.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat #count_tokens with a provider without token counting raises a clear error","notes":"Checks the error on the default protocol."}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no image generation endpoint, and models list text as their only output. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s770"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no image editing endpoint, and models list text as their only output. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s770"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no video generation endpoint, and models list text as their only output.","sources":["s164","s770"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no speech endpoint, and models list text as their only output. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s770"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no transcription endpoint. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no transcription endpoint, streaming or otherwise.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no transcription endpoint, so it offers no speaker identification.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no moderation endpoint. A moderation key in an example response and user_id content safety handling are not a moderation operation. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s166","s775"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no OCR endpoint. Reading text from a screenshot through chat is image input.","sources":["s164","s768"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no embeddings endpoint. The Chat Completions protocol carries this operation, so a RubyLLM call reaches an endpoint DeepSeek does not list.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no embeddings endpoint.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no reranking endpoint. The rerank_score field on Search results is a search ranking value.","sources":["s164","s776"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload and UploadedFile.find support DeepSeek image files and expires_in, with purpose defaulting to user_data, the only documented value. RubyLLM leaves file types and sizes to DeepSeek, which accepts JPEG, PNG, GIF and WebP up to 64 MiB.","sources":["s168","s777","s778"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/deepseek/files.rb","method":"render_upload_fields"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/files.rb","method":"parse_file_response"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-10-25","spec":"spec/ruby_llm/protocols/deepseek/files_spec.rb","example":"RubyLLM::Protocols::DeepSeek::Files uploads and retrieves an image with an expiry","cassette":"spec/fixtures/vcr_cassettes/protocols_deepseek_files_uploads_and_retrieves_an_image_with_an_expiry.yml","notes":"Uploads, finds and checks the expiry of an image. The request names no model."}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"The Files API documents upload, list, retrieve metadata and delete, and no content download. RubyLLM raises an explicit error for download and marks files as not downloadable.","sources":["s168","s164"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/deepseek/files.rb","method":"download"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/deepseek/files_spec.rb","example":"RubyLLM::Protocols::DeepSeek::Files reports that stored images cannot be downloaded","notes":"Checks the error and the downloadable flag."}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no batch endpoint. Off-peak pricing is a lower rate on normal requests, not a batch API. RubyLLM raises for batches on this provider.","sources":["s164","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch .submit rejects providers without batch support","notes":"Checks the error."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory has no batch endpoint. Off-peak pricing is a lower rate on normal requests, not a batch API. RubyLLM raises for batches on this provider.","sources":["s164","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document advisor tool. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Perplexity Agent API, which is not an API of this provider.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document agent skills. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document asynchronous research. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API is stateless and does not support background.","sources":["s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document browser use. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document classification. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API lists computer_use among built-in tools it ignores.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document context editing. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document contextual embeddings. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document dedicated transcription api. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document dubbing. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API does not support service_tier, and no reference page documents a faster tier. Peak and off-peak rates change price, not speed.","sources":["s165","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"DeepSeek documents listing and deleting image files. RubyLLM offers upload, find and download on UploadedFile, with no list or delete.","sources":["s168","s779","s780"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/uploaded_file.rb","method":"find"},{"version":"v2","path":"lib/ruby_llm/protocols/deepseek/files.rb"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document file search store management. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"fill_in_middle":{"v1":"missing","v2":"missing","offered":true,"notes":"DeepSeek documents FIM completion in beta at /completions on its /beta base URL, with prompt and suffix, in non-thinking mode. RubyLLM has no completions operation for it.","sources":["s781","s782","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document fine-tuning. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google Maps grounding, which is not an API of this provider.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document guardrail configuration. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API is stateless and does not support conversation, store or previous_response_id.","sources":["s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document image batches. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Interactions API, which is not an API of this provider.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"JSON mode uses raw options: with_provider_options(response_format: { type: \"json_object\" }) on Chat Completions, or text: { format: { type: \"json_object\" } } on Responses. with_schema on Responses is the dedicated JSON Schema API, counted separately.","sources":["s783","s766","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"deepseek-v4-flash","spec":"spec/ruby_llm/chat_request_options_spec.rb","example":"RubyLLM::Chat with params deepseek/deepseek-v4-flash supports response_format param","cassette":"spec/fixtures/vcr_cassettes/chat_with_params_deepseek_deepseek-v4-flash_supports_response_format_param.yml","notes":"Sends response_format json_object on Chat Completions and parses the JSON answer."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names Amazon Bedrock Mantle, which is not an API of this provider.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no compaction endpoint, and Responses does not support context_management. RubyLLM raises for chat.compact on this provider.","sources":["s165","s164"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat does not enable manual compaction for other Responses dialects","notes":"Checks the error."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document music generation. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document other batch operations. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Omni video endpoint, which is not an API of this provider.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"DeepSeek serves only its own models and hosts no partner models. Its OpenAI and Anthropic formats are ways to reach those same models.","sources":["s767","s769"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"prefix_completion":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"DeepSeek documents Chat Prefix Completion in beta. It needs the /beta base URL through deepseek_api_base and a final assistant message with prefix: true. RubyLLM has no prefix option, so pass the whole message list through with_provider_options(messages: [...]), which replaces the rendered transcript. No API recording covers it.","sources":["s784","s767"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v1","path":"lib/ruby_llm/chat.rb","method":"with_tools"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb","method":"api_base"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document programmatic tool calling. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API does not support prompt_cache_key or prompt_cache_retention, and caching is managed automatically. The user_id parameter isolates caches between end users, and RubyLLM sends it through with_end_user.","sources":["s165","s773","s775"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/chat.rb","method":"apply_end_user"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document provider-managed fallback. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document realtime sessions. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document response caching. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"The provider registers a DeepSeek Responses dialect. Select it with protocol: :responses, as Chat Completions is the default. It handles reasoning_text, images, schemas and function tools. The API is stateless, and RubyLLM sends the local transcript.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"parse_reasoning_summary, build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-10-17","model":"deepseek-v4-flash","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses sends the phase of a streamed answer back on the next turn","cassette":"spec/fixtures/vcr_cassettes/providers_deepseek_responses_sends_the_phase_of_a_streamed_answer_back_on_the_next_turn.yml","notes":"Two streamed turns on /responses."}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document responses batches. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"search_api":{"v1":"na","v2":"missing","offered":true,"notes":"DeepSeek documents a standalone Search API at POST /v1/tools/search, billed per successful request. RubyLLM has no operation for it, and it sits outside the chat payload, so raw options cannot reach it.","sources":["s776","s785","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"server_apply_patch":{"v1":"missing","v2":"partial","offered":true,"notes":"The :apply_patch alias with protocol: :responses sends the custom tool definition, the only custom tool DeepSeek accepts. RubyLLM records the returned custom_tool_call as a ServerToolCall instead of running a local tool, and has no way to send custom_tool_call_output back, so the loop is left to the application.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek/responses.rb","method":"SERVER_TOOL_ALIASES, render_tool_items"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/deepseek/responses_spec.rb","example":"RubyLLM::Providers::DeepSeek::Responses server tools keeps the patch tool alias","notes":"Checks the rendered tool definition."}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Responses reference allows only function, custom and web_search tool types and ignores other built-in tools. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API does not run tool search: with 13 deferred functions and a tool_search tool, deepseek-v4-flash called the function directly in every run and no tool_search_call appeared. The API also accepts unknown tool types silently, so a tool_search entry alone does not mean it runs. The tools array documents only function, custom and web_search types.","sources":["s166","s165"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_tool_search"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"deepseek-v4-flash","notes":"Three requests with deferred functions and a tool_search tool returned function calls and no tool_search_call. A nonsense tool type was also accepted."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document x search tool. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The Responses reference allows only function, custom and web_search tool types, and ignores computer_use and other built-in tools.","sources":["s165","s166"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document sound effects. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document streaming speech generation. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document task budgets. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document text analysis api. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document vector store management. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document video batches. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document video editing. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document video extension. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document voice cloning. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document voice conversion. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"web_fetch_api":{"v1":"na","v2":"missing","offered":true,"notes":"DeepSeek documents a standalone Fetch API at POST /v1/tools/fetch that returns a page as Markdown, billed per successful request. RubyLLM has no operation for it, and it sits outside the chat payload, so raw options cannot reach it.","sources":["s776","s786","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The DeepSeek API reference does not document responses over websocket. The inventory covers Chat Completions, Responses, the Anthropic-format Messages, FIM completion, Files, Search, Fetch, models and balance.","sources":["s164","s774"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/deepseek.rb"},{"version":"v2","path":"lib/ruby_llm/providers/deepseek.rb"}]}},"gaps":[{"notes":"File listing and deletion: DeepSeek documents file listing and deletion, and RubyLLM has no list or delete operation.","sources":["s168","s779","s780"]},{"notes":"Fill-in-the-middle completion: DeepSeek documents a beta FIM Completion API on its /beta base URL, and RubyLLM has no completions operation.","sources":["s781","s782"]},{"notes":"Standalone search API, Standalone web fetch API: DeepSeek documents POST /v1/tools/search and POST /v1/tools/fetch, and RubyLLM wraps neither.","sources":["s776","s785","s786"]},{"notes":"Apply patch tool: The alias sends the custom apply_patch definition, but custom_tool_call output is recorded as a ServerToolCall, with no local tool loop and no custom_tool_call_output.","sources":["s165"]},{"notes":"JSON object mode: JSON mode uses raw response_format on Chat Completions or raw text.format on Responses.","sources":["s783","s166"]},{"notes":"Prefix completion: Chat Prefix Completion needs the /beta base URL and a raw message list with prefix: true. RubyLLM has no prefix option.","sources":["s784"]},{"notes":"Image input: Uploaded file references work only on Responses. Chat Completions raises UnsupportedAttachmentError, although DeepSeek documents a file content block there.","sources":["s768","s168"]},{"notes":"Thinking controls: On Responses, with_thinking(false) sends no reasoning field and thinking stays on by default. Use effort: :none.","sources":["s771","s165"]}]},{"id":"cohere","name":"Cohere","notes":"Audited: the public Cohere inference and dataset APIs, with Chat v2 as the RubyLLM route and v1 endpoints where Cohere still documents them. North, Model Vault, the OpenAI-compatible Compatibility API and third-party or private deployments are outside this scope.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM sends conversations to /v2/chat from the local transcript. Sampling options without a shared name, such as p, k, seed and priority, go through with_provider_options.","sources":["s84","s787"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"render_payload, parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality cohere/command-a-03-2025 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_cohere_command-a-03-2025_can_have_a_basic_conversation.yml","notes":"Answers and reports token usage."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"Parses the v2 event stream for text, thinking, tool calls, citations and final usage. A stream that ends with finish_reason ERROR raises a ServerError.","sources":["s84","s788"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/streaming.rb","method":"build_chunk, parse_content_delta, stream_usage"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses cohere/command-a-03-2025 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_cohere_command-a-03-2025_supports_streaming_responses.yml","notes":"Streams chunks through the public chat API."}]},"structured_output":{"v1":"missing","v2":"native","offered":true,"notes":"with_schema renders response_format type json_object with a bare json_schema. Cohere documents this on Command R, Command R+ and newer models. Chat batches reject response_format before upload.","sources":["s84","s789"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"build_response_format"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat#with_schema with cohere/command-a-03-2025 accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_cohere_command-a-03-2025_accepts_a_json_schema_and_returns_structured_output.yml","notes":"Returns parsed JSON that matches the schema."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Chat sends image_url blocks on Cohere vision models, such as Command A Vision and Command A+. A remote URL without a known image extension is inlined as a data URI, because Cohere resolves remote images by extension.","sources":["s84","s790","s326"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/media.rb","method":"render_image, image_source"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-plus-05-2026","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models cohere/command-a-plus-05-2026 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_cohere_command-a-plus-05-2026_can_understand_local_images.yml","notes":"Describes a local image."}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 user content documents text and image_url blocks only, so RubyLLM raises UnsupportedAttachmentError for a PDF. The Parse API reference states PDF and file URL inputs are not yet supported. Mixed PDF page input on Embed models is an embedding input, not a chat attachment.","sources":["s84","s85"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/media.rb","method":"supported_attachment?"}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 message content documents text and images, with no audio block. Audio goes through the separate transcription endpoint.","sources":["s84","s90"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/media.rb","method":"supported_attachment?"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 message content documents text and images, with no video block.","sources":["s84","s326"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/media.rb","method":"supported_attachment?"}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"Returns thinking content blocks on reasoning models, such as Command A Reasoning, in full and streaming responses. On a tool-calling turn without thinking blocks, tool_plan is exposed as thinking and stays out of later requests.","sources":["s87","s84"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"parse_thinking, render_assistant_message"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/streaming.rb","method":"parse_thinking_delta"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-reasoning-08-2025","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking cohere/command-a-reasoning-08-2025 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_cohere_command-a-reasoning-08-2025_returns_thinking_when_available.yml","notes":"Returns thinking text from a reasoning model."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking renders thinking type enabled or disabled, with token_budget from budget:. The wire format has no effort field, so only the toggle and the budget reach Cohere.","sources":["s87","s84"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"add_thinking"}]},"citations":{"v1":"missing","v2":"native","offered":true,"notes":"with_citations sends text attachments as top-level documents and parses typed citations, including citations of tool results returned as search results and streamed citation events. citation_options goes through with_provider_options.","sources":["s84","s791"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"parse_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/media.rb","method":"render_documents"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/tools.rb","method":"document_blocks"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/streaming.rb","method":"parse_citations_delta"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-plus-05-2026","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with cohere/command-a-plus-05-2026 cites text documents in responses","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_cohere_command-a-plus-05-2026_cites_text_documents_in_responses.yml","notes":"Returns citations of an attached text document."},{"status":"passed","recorded_on":"2026-10-11","model":"command-a-plus-05-2026","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with cohere/command-a-plus-05-2026 streams citations","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_cohere_command-a-plus-05-2026_streams_citations.yml","notes":"Returns citations from a stream."}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"Sends client function tools and replays their results. strict_tools has no shared name and goes through with_provider_options.","sources":["s84","s792"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/tools.rb","method":"function_for, render_tool_result, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling cohere/command-a-03-2025 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_cohere_command-a-03-2025_can_use_tools.yml","notes":"Runs a local tool and answers from its result."}]},"tool_choice":{"v1":"missing","v2":"native","offered":true,"notes":"Supports the choices Cohere exposes: auto by omission, REQUIRED and NONE. Cohere cannot force a named tool, so RubyLLM raises ArgumentError for one. Documented for Command R7B and newer models.","sources":["s84"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control cohere/command-a-03-2025 respects choice: :none","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_cohere_command-a-03-2025_respects_choice_none.yml","notes":"Sends NONE and gets a direct answer."}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"Parses and runs every tool call in one assistant response, in full and streaming modes.","sources":["s793","s84"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/tools.rb","method":"parse_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/streaming.rb","method":"stream_tool_call_start"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling cohere/command-a-03-2025 can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_cohere_command-a-03-2025_can_use_parallel_tool_calls.yml","notes":"The recorded response carries two tool calls."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 documents no option that limits parallel calls, and the Compatibility API lists parallel_tool_calls as unsupported. The calls: option on with_tool_options limits calls in RubyLLM and sends nothing to Cohere.","sources":["s84","s794"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"add_tools"}]},"server_web_search":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 tools are client-defined functions. The web-search connector belongs to the v1 connectors parameter, which Cohere deprecated effective 2025-09-15 and does not carry into v2. Its migration guide replaces it with a client tool.","sources":["s84","s170","s795"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 tools are client-defined functions. No hosted fetch tool is documented.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 tools are client-defined functions. No hosted code execution tool is documented.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 tools are client-defined functions. Documents are supplied by the caller, and the managed connectors were deprecated effective 2025-09-15.","sources":["s84","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 tools are client-defined functions. No remote MCP tool type is documented. Cohere's documentation MCP server and North's connectors are not inference tools.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"Cohere caches on its own, with nothing to request. Chat v2 reports hits as usage.cached_tokens and recorded responses show nonzero values, but the docs do not describe the cache or its pricing. A bare with_caching is accepted and options raise ArgumentError. RubyLLM counts billed_units, so cache_read stays nil: cached_tokens includes a preamble Cohere does not bill. The field stays readable through raw.","sources":["s84"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"reject_caching_options, usage_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality cohere/command-a-03-2025 returns raw responses","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_cohere_command-a-03-2025_returns_raw_responses.yml","notes":"The recorded response reports cached_tokens of 448. The example does not assert it."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/cohere/chat_spec.rb","example":"RubyLLM::Protocols::Cohere::Chat#parse_completion_body when the reply hit the prompt cache reports the billed units without the cache hits Cohere does not bill","notes":"Usage stays on billed units when cached_tokens is present."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 documents no such request option or endpoint; cached_tokens is only a response field. Cache markers cannot be sent.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 documents no such request option or endpoint; cached_tokens is only a response field. No cache resource API is listed.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2, the API RubyLLM speaks, does not support prompt_truncation. Only the older v1 Chat offers it.","sources":["s796","s795","s797"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/chat.rb","method":"completion_url"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.tokenize calls /v1/tokenize for exact plain-text token IDs. Chat#count_tokens is not implemented for Cohere, which documents no whole-request counter. Tokenizing records no generation usage.","sources":["s86","s798"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/tokenization.rb","method":"tokenization_url, render_tokenization_payload, parse_tokenization_response"},{"version":"v2","path":"lib/ruby_llm/tokenization.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/tokenization_spec.rb","example":"RubyLLM::Tokenization tokenizes text with cohere through the public API","cassette":"spec/fixtures/vcr_cassettes/tokenization_tokenizes_text_with_cohere_through_the_public_api.yml","notes":"Returns integer token IDs for command-a-03-2025."}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"No image generation operation is documented in the Cohere API inventory. Parse, Embed, Transcribe and Command vision input are separate operations.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"No image editing operation is documented in the Cohere API inventory. Parse, Embed, Transcribe and Command vision input are separate operations.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"No video generation operation is documented in the Cohere API inventory. Parse, Embed, Transcribe and Command vision input are separate operations.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"No speech generation operation is documented in the Cohere API inventory. Parse, Embed, Transcribe and Command vision input are separate operations.","sources":["s169","s799","s800","s170","s326"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts to /v2/audio/transcriptions with language and temperature. Cohere requires language, so RubyLLM sends en when none is given. The endpoint returns text only and takes flac, mp3, mpeg, mpga, ogg and wav files. format, prompt, speaker names and timestamps raise before the request.","sources":["s90","s801"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/transcription.rb","method":"render_transcription_payload, parse_transcription_response, rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality cohere/cohere-transcribe-03-2026 can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_cohere_cohere-transcribe-03-2026_can_transcribe_audio.yml","notes":"Transcribes an audio file."}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The transcription endpoint documents no streaming, and the Compatibility API lists stream as unsupported for transcriptions.","sources":["s90","s794"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/transcription.rb","method":"render_transcription_payload"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The model pages state Cohere Transcribe has neither timestamps nor speaker diarization.","sources":["s90","s801"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/transcription.rb","method":"parse_transcription_response"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The API reference lists no moderation endpoint. An old release note announces a Moderate beta, with no reference page today. safety_mode changes chat behavior and is a separate row.","sources":["s169","s799","s800","s170","s802"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"ocr":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.ocr calls /v2/parse with one image URL or inline image. Default output is Markdown, and provider_options output_format: blocks also works. Pages normalize to OCR::Page and keep the original page hash, image and table metadata, and billed pages. The API reference accepts images only and says PDF and file URL inputs are not yet supported, although the model page lists PDF and PPT, so PDFs raise before the request. pages may be omitted or [0]; other selections raise.","sources":["s85","s803","s325","s327"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/ocr.rb","method":"render_ocr_payload, parse_ocr_page"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/models.rb","method":"capabilities_from"},{"version":"v2","path":"lib/ruby_llm/ocr.rb","method":"pages"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"parse-v5.0","spec":"spec/ruby_llm/protocols/cohere/ocr_spec.rb","example":"RubyLLM::Protocols::Cohere::OCR parses an image through the public OCR API with markdown output","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_ocr_parses_an_image_through_the_public_ocr_api_with_markdown_output.yml","notes":"Markdown output with images and billed pages."},{"status":"passed","recorded_on":"2026-10-11","model":"parse-v5.0","spec":"spec/ruby_llm/protocols/cohere/ocr_spec.rb","example":"RubyLLM::Protocols::Cohere::OCR parses an image through the public OCR API with blocks output","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_ocr_parses_an_image_through_the_public_ocr_api_with_blocks_output.yml","notes":"Blocks output normalized to the same page object."}]},"embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.embed calls /v2/embed for Embed v3, v4 and v5 models and returns float vectors. task_type sets input_type, which defaults to search_document. dimensions sets output_dimension on Embed v4 and newer. int8, uint8 and binary types are not parsed.","sources":["s804","s88"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/embeddings.rb","method":"render_embedding_payload, parse_embedding_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"embed-v4.0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality cohere/embed-v4.0 can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_cohere_embed-v4_0_can_handle_a_single_text.yml","notes":"Returns a float vector."},{"status":"passed","recorded_on":"2026-10-11","model":"embed-v4.0","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality cohere/embed-v4.0 can handle a single text with custom dimensions","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_cohere_embed-v4_0_can_handle_a_single_text_with_custom_dimensions.yml","notes":"Honors the requested dimensions."}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"Embed v4 and v5 take mixed text and images through the shared embedding API using inputs. Embed v3 takes one image without text using images, which RubyLLM selects for embed-english-v3.0 and embed-multilingual-v3.0; the light v3 models take the inputs path. Azure-hosted Cohere deployments reuse the same serialization. The recorded v3 request returned a 1024-dimensional vector and billed one image.","sources":["s804","s88","s805"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/embeddings.rb","method":"separate_image_embeddings?, image_embedding_inputs, embedding_inputs"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"embed-english-v3.0","spec":"spec/ruby_llm/protocols/cohere/embeddings_spec.rb","example":"RubyLLM::Protocols::Cohere::Embeddings embeds an image with native Cohere Embed v3","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_embeddings_embeds_an_image_with_native_cohere_embed_v3.yml","notes":"Embeds one image with embed-english-v3.0."}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.rerank calls /v2/rerank, resolves ranked text from the documents sent and reports search units. relevant_spans (Context Focus on Rerank 5) and max_tokens_per_doc go through provider_options; returned spans are readable only through raw.","sources":["s89","s806"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/rerank.rb","method":"render_rerank_payload, parse_rerank_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"rerank-v3.5","spec":"spec/ruby_llm/rerank_spec.rb","example":"RubyLLM::Rerank reranking with cohere/rerank-v3.5 orders documents by relevance and resolves them from the request","cassette":"spec/fixtures/vcr_cassettes/rerank_reranking_with_cohere_rerank-v3_5_orders_documents_by_relevance_and_resolves_them_from_the_request.yml","notes":"Orders two documents and maps indexes back to them."}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload creates Cohere datasets. purpose selects the dataset type, provider_options sets name and other query options, and the validation status is exposed. Datasets feed batches and expire after 30 days; they are not chat attachments.","sources":["s369","s370","s371"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/datasets.rb","method":"upload, wait_for_validation"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/protocols/cohere/datasets_spec.rb","example":"RubyLLM::Protocols::Cohere::Datasets uploads validates retrieves and downloads a Cohere dataset","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_datasets_uploads_validates_retrieves_and_downloads_a_cohere_dataset.yml","notes":"Uploads, validates, finds and downloads one embed-input dataset."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"Original uploaded bytes download unchanged by default. Generated Avro dataset parts decode in provider part order to JSONL through the optional avro gem. Signed URLs are fetched without API credentials.","sources":["s370","s371"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/datasets.rb","method":"download, records"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/protocols/cohere/datasets_spec.rb","example":"RubyLLM::Protocols::Cohere::Datasets uploads validates retrieves and downloads a Cohere dataset","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_datasets_uploads_validates_retrieves_and_downloads_a_cohere_dataset.yml","notes":"Uploads, validates, finds and downloads one embed-input dataset."}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"Create, find, cancel and results use /v2/batches with batch-chat-v2-input datasets. Results keep IDs, tool calls, model and per-request usage, and failed rows become failed slots. Structured output, forced tool choice and retrieval documents raise before upload. Cancel posts to /v2/batches/{id}/cancel while the reference documents {id}:cancel, and that route has no recording.","sources":["s173","s174","s807","s369"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/batches.rb","method":"create_batch, find_batch, cancel_batch, batch_results"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/batch_requests.rb","method":"render_batch_request, render_batch_chat"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"command-a-03-2025","spec":"spec/ruby_llm/protocols/cohere/batches_spec.rb","example":"RubyLLM::Protocols::Cohere::Batches collects completed Cohere chat responses with returned IDs and token usage","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_batches_collects_completed_cohere_chat_responses_with_returned_ids_and_token_usage.yml","notes":"Runs a two-request batch to completion."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/cohere/batches_spec.rb","example":"RubyLLM::Protocols::Cohere::Batches collects typed messages and individual failures with provider usage and returned tool calls","notes":"Covers failed rows and tool calls against stubbed data."}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"Uses /v2/batches with batch-embed-v2-input datasets, recorded on embed-v4.0, separately from the older Embed Jobs API that is limited to v3 models and has no RubyLLM route. Scalar and array shapes survive Batch.find and out-of-order output. dimensions is sent as output_dimension and left to Cohere to validate. Cancel shares the undocumented route noted for chat batches.","sources":["s173","s174","s369","s91"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/cohere/batches.rb","method":"parse_batch_result"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/batch_requests.rb","method":"batch_dataset_type, render_batch_request"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"embed-v4.0","spec":"spec/ruby_llm/protocols/cohere/batches_spec.rb","example":"RubyLLM::Protocols::Cohere::Batches collects completed Cohere embedding batches with scalar and array results","cassette":"spec/fixtures/vcr_cassettes/protocols_cohere_batches_collects_completed_cohere_embedding_batches_with_scalar_and_array_results.yml","notes":"Runs scalar and array embedding requests to completion."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/cohere/batches_spec.rb","example":"RubyLLM::Protocols::Cohere::Batches leaves embedding dimensions to Cohere","notes":"output_dimension passes through the batch body."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"No advisor tool operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Perplexity Agent API; it is not an API of this provider.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"No agent skills operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"No asynchronous research operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"No background responses jobs operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"No browser use operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"Cohere deprecated /v1/classify effective 2025-09-15, and its lifecycle definition closes deprecated endpoints to new customers. No shutdown date is published, and the legacy reference page stays online.","sources":["s170","s808"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"No computer use operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"No context editing operation is documented in the Cohere API inventory. v1 prompt_truncation is covered by Context compaction.","sources":["s169","s799","s800","s170","s796"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"No contextual embeddings operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM implements the dedicated /v2/audio/transcriptions endpoint. This row overlaps Transcription.","sources":["s90"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"No dubbing operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"No fast tier is documented. The priority option on chat and rerank ranks requests under load, and its default of 0 is already the highest.","sources":["s84","s89"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Cohere lists and deletes datasets. RubyLLM has upload, find and download only, and raises for file listing. These datasets feed batches and embed jobs, not conversations.","sources":["s171","s172"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"list_uris"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/datasets.rb"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No file search store management operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"No fill-in-the-middle operation is documented. The legacy /v1/generate endpoint takes a prompt with no suffix parameter and was deprecated effective 2025-09-15.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"Cohere retired fine-tuning through the dashboard and API effective 2025-09-15, and earlier fine-tuned models are inaccessible. Legacy reference pages stay online.","sources":["s170","s808"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Maps grounding; it is not an API of this provider.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"guardrail_configuration":{"v1":"na","v2":"passthrough","offered":true,"notes":"with_provider_options(safety_mode: \"STRICT\") selects CONTEXTUAL, STRICT or OFF on Chat v2. The value is merged into the request. Cohere ignores it with tools or documents, and OFF is unavailable on Command R7B and newer. Chat batches reject it.","sources":["s84","s809"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"hosted_conversations":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"v1 Chat takes conversation_id to create or resume a conversation Cohere persists. Chat v2 does not support it. RubyLLM builds requests from the local transcript.","sources":["s796","s795","s797"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"No image batches operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Interactions API; it is not an API of this provider.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"json_mode":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Chat v2 accepts response_format type json_object without a schema. with_provider_options merges that value into the request; with_schema covers schema output separately.","sources":["s84","s789"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Amazon Bedrock Mantle; it is not an API of this provider.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"No compaction operation is documented in the Cohere API inventory. v1 prompt_truncation is automatic and covered by Context compaction.","sources":["s169","s799","s800","s170","s796"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"No music generation operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Cohere batch dataset types cover chat and embedding requests only. Both have their own rows.","sources":["s369","s173"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/cohere/batch_requests.rb","method":"batch_dataset_type"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row specifically names Google Omni video endpoint; it is not an API of this provider.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The public API serves Cohere models over its own wire format and an OpenAI-compatible Compatibility API. Open models run on a customer Model Vault endpoint, a separate deployment outside this API.","sources":["s794","s800","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 documents no assistant prefill or prefix option. The legacy /v1/generate endpoint takes a prompt with no suffix parameter and was deprecated effective 2025-09-15.","sources":["s84","s170","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"No programmatic tool calling operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"Chat v2 documents no such request option or endpoint; cached_tokens is only a response field.","sources":["s84","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"No provider-managed fallback operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"No realtime session API is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"No response caching operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"No Responses endpoint is documented. The Compatibility API covers chat completions, embeddings and audio transcriptions.","sources":["s794","s169"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"No responses batches operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"No standalone search api operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"No apply patch tool operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"No image generation tool operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"No tool search operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"No x search tool operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"No shell and computer tools operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"No sound effects operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"No speech generation operation is documented in the Cohere API inventory. Parse, Embed, Transcribe and Command vision input are separate operations.","sources":["s169","s799","s800","s170","s326"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"No task budgets operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"No text analysis API is active. The legacy /v1/summarize endpoint and model were deprecated effective 2025-09-15, in favor of Chat.","sources":["s170","s808"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No vector store management operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"No video batches operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"No video editing operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"No video extension operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"No voice cloning operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"No voice conversion operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"No standalone web fetch api operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"No responses over websocket operation is documented in the Cohere API inventory.","sources":["s169","s799","s800","s170"],"code":[{"version":"v1","path":"lib/ruby_llm.rb"},{"version":"v2","path":"lib/ruby_llm/providers/cohere.rb"}]}},"gaps":[{"notes":"File listing and deletion: Cohere lists and deletes datasets. RubyLLM has no dataset listing or deletion; these resources feed batches, not conversations.","sources":["s171","s172"]},{"notes":"JSON object mode, Guardrail configuration: response_format type json_object without a schema and safety_mode need with_provider_options.","sources":["s84","s809"]},{"notes":"Chat batches, Embedding batches: cancel posts to /v2/batches/{id}/cancel, while the reference documents /v2/batches/{id}:cancel. The route has no recording. The older Embed Jobs API for v3 models has no RubyLLM route.","sources":["s807","s91"]},{"notes":"Reranking: relevant_spans on Rerank 5 goes through provider_options, and returned spans are readable only through raw.","sources":["s89"]},{"notes":"Multimodal embeddings: light Embed v3 models take the inputs path rather than images, and no recording covers them. Non-float embedding types are not parsed.","sources":["s804","s88"]}]},{"id":"mistral","name":"Mistral","notes":"Covers Mistral inference, hosted tools, files, batches and audio on api.mistral.ai. RubyLLM uses Chat Completions with multi-completion hosted tools by default and an explicit stateless Conversations protocol. Stored conversations, the Agents SDK harness, Vibe, and agent, library and skill resources stay separate.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.chat uses /v1/chat/completions by default. protocol: :conversations selects stateless Conversations for hosted tools. Model support varies; the API rejects what a model cannot do.","sources":["s92","s810"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions.rb","method":"ChatCompletions"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality mistral/mistral-small-latest can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_mistral_mistral-small-latest_can_have_a_basic_conversation.yml","notes":"Basic request and response on Chat Completions."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Chat Completions and Conversations both stream over server-sent events. Requests with hosted tools on Chat Completions use the multi-completion stream.","sources":["s92","s810","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/multi_completion.rb","method":"stream_response"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses mistral/mistral-small-latest supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_mistral_mistral-small-latest_supports_streaming_responses.yml","notes":"Streamed text deltas on Chat Completions."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema renders response_format with json_schema on Chat Completions, and inside completion_args on Conversations. Model support varies; the API rejects what a model cannot do.","sources":["s811","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"completion_args"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/capabilities.rb","method":"STRUCTURED_OUTPUT_MODELS"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat with schema with mistral/mistral-small-latest accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_mistral_mistral-small-latest_accepts_a_json_schema_and_returns_structured_output.yml","notes":"The recorded request carries response_format with json_schema."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as image_url content parts, by URL or data URI, on vision-capable models.","sources":["s812","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/media.rb","method":"render_image"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"pixtral-12b","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models mistral/pixtral-12b can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_mistral_pixtral-12b_can_understand_local_images.yml","notes":"Local image sent as a data URI."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDF and document attachments render as document_url parts, by URL or data URI, on models with document support. Uploaded file IDs cannot be attached: the Mistral renderer has no file part. A DOCX recording covers the document_url path; no recording covers a PDF in chat.","sources":["s813","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/media.rb","method":"render_document"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/media.rb","method":"supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat document models mistral/mistral-small-latest understands DOCX documents","cassette":"spec/fixtures/vcr_cassettes/chat_document_models_mistral_mistral-small-latest_understands_docx_documents.yml","notes":"Covers document_url with a DOCX file, the same content part a PDF uses."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Audio attachments render as input_audio parts on Voxtral chat models.","sources":["s814","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/media.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_audio"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-small-latest","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models mistral/voxtral-small-latest can understand audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_mistral_voxtral-small-latest_can_understand_audio.yml","notes":"Local audio sent to a Voxtral chat model."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The chat content schema documents text, images, documents, references, files, thinking and audio, with no video part. The vision guide and current model cards list no video input. Extracted image frames and Vibe application behavior are separate.","sources":["s815","s812","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Thinking chunks parse into Message#thinking on sync and streamed responses, and replay as thinking content blocks across turns.","sources":["s816","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"build_thinking_blocks"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_content_and_thinking"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat Mistral hybrid reasoning separates thinking from final content","cassette":"spec/fixtures/vcr_cassettes/chat_mistral_hybrid_reasoning_separates_thinking_from_final_content.yml","notes":"The recorded response carries a thinking chunk parsed apart from the answer."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort, inside completion_args on Conversations. Accepted values differ by model family. A thinking budget raises ArgumentError because Mistral has no budget control.","sources":["s816","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v2","path":"lib/ruby_llm/thinking.rb","method":"Thinking"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat Mistral hybrid reasoning separates thinking from final content","cassette":"spec/fixtures/vcr_cassettes/chat_mistral_hybrid_reasoning_separates_thinking_from_final_content.yml","notes":"The recorded request sends reasoning_effort high."}]},"citations":{"v1":"missing","v2":"native","offered":true,"notes":"Hosted tool results become Citation values: tool_reference chunks on Conversations and multi-completion Chat Completions, including streamed results. Reference chunks with reference_ids, returned when your own tool supplies numbered sources, become Citation values too on Chat Completions: source_index is the key in that tool's result, and non-streamed citations carry the offset of the reference in the answer, while streamed ones carry none. Conversations reference chunks are not checked and are not parsed.","sources":["s817","s301","s302","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"extract_citations, parse_chunk_citations, reference_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/content.rb","method":"parse_conversation_citation"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/multi_completion.rb","method":"multi_content"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"mistral-large-latest","spec":"spec/ruby_llm/providers/mistral/reference_citations_spec.rb","example":"Mistral reference citations turns reference chunks into citations pointing at the tool result keys","cassette":"spec/fixtures/vcr_cassettes/mistral_reference_citations_turns_reference_chunks_into_citations_pointing_at_the_tool_result_keys.yml","notes":"reference chunks with reference_ids after a tool result, as Citation values with offsets."},{"status":"passed","recorded_on":"2026-11-10","model":"mistral-large-latest","spec":"spec/ruby_llm/providers/mistral/reference_citations_spec.rb","example":"Mistral reference citations collects reference chunks from a stream","cassette":"spec/fixtures/vcr_cassettes/mistral_reference_citations_collects_reference_chunks_from_a_stream.yml","notes":"Streamed reference chunks become Citation values."},{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations_live_spec.rb","example":"searches the web with citations and replays hosted results in a stateless conversation","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_searches_the_web_with_citations_and_replays_hosted_results_in_a_stateless_conversation.yml","notes":"Hosted web search returned citations, and the follow-up replayed the results without stored state."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools render on Chat Completions and Conversations. Function results are replayed from the local transcript.","sources":["s818","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"render_conversation_message"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling mistral/mistral-small-latest can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_mistral_mistral-small-latest_can_use_tools.yml","notes":"Tool call and result round trip."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps :required to any, or to the single named tool when only one is present. Conversations accepts :auto, :none and :required, and raises on a named tool.","sources":["s818","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"normalize_required_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"normalize_conversation_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control mistral/mistral-small-latest respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_mistral_mistral-small-latest_respects_specific_tool_choice.yml","notes":"The recorded request names the function in tool_choice."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool calls in one response parse into separate ToolCall values.","sources":["s818","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling mistral/mistral-small-latest can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_mistral_mistral-small-latest_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"The recorded response returns more than one tool call."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) sends parallel_tool_calls false on Chat Completions. Conversations has no such control and raises ArgumentError.","sources":["s818","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control mistral/mistral-small-latest respects calls: :one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_mistral_mistral-small-latest_respects_calls_one_for_sequential_execution.yml","notes":"The recorded request sends parallel_tool_calls false."}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) on the explicit :conversations protocol. Completed hosted calls stay ServerToolCall values with raw results and usage, and follow-ups replay them statelessly. The web_search_premium variant has no alias. Mistral says the Agents SDK replaces the Agents and Conversations API and calls Agents API v1 about to be deprecated, with no date.","sources":["s182","s301","s179","s819"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"parse_conversation_steps"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations_live_spec.rb","example":"searches the web with citations and replays hosted results in a stateless conversation","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_searches_the_web_with_citations_and_replays_hosted_results_in_a_stateless_conversation.yml","notes":"Hosted web search returned citations, and the follow-up replayed the results without stored state."}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_fetch) enables the Conversations web_search tool, whose open_url step fetches a page. The web search guide names open_url under domain filtering, without a dedicated section.","sources":["s182","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"parse_conversation_steps"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations_live_spec.rb","example":"fetches a public page through the web fetch alias","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_fetches_a_public_page_through_the_web_fetch_alias.yml","notes":"The recorded response shows the open_url step fetching the requested page."}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:code_execution) maps to code_interpreter on Conversations, including streamed results, execution records, usage and stateless continuation.","sources":["s820","s301","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"replay_conversation_execution"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations_live_spec.rb","example":"streams hosted Python execution with complete tool history and usage","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_streams_hosted_python_execution_with_complete_tool_history_and_usage.yml","notes":"Hosted Python ran in a streamed request with its history and usage retained."}]},"server_file_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(file_search: {library_ids: [...]}) maps to document_library on Conversations. It uses existing indexed libraries; creating and managing them is a separate row.","sources":["s360","s301","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"parse_conversation_steps"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations_live_spec.rb","example":"searches an uploaded document through the file search alias","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_searches_an_uploaded_document_through_the_file_search_alias.yml","notes":"The spec prepares a library, then the alias searches it through document_library."}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"The mcp alias maps connector_id to connector tools on Chat Completions and on stateless Conversations. Connectors are registered MCP servers in public preview. Completed results replay from local history. Confirmations need stored conversations and raise. No recording covers remote execution.","sources":["s821","s822","s362","s302","s183"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/multi_completion.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"requires_confirmation"}],"verification":[{"status":"unavailable","checked_on":"2026-09-07","notes":"A fresh request to create a private read-only Microsoft Learn MCP connector returned HTTP 403: this API key lacks personal_and_shared scope. No connector created, permissions changed, or existing business tools invoked. Remote connector execution is not live-verified."},{"status":"unit","spec":"spec/ruby_llm/protocols/mistral/conversations_spec.rb","notes":"Stateless hosted tool rendering and result replay are covered. Storage and confirmation options are rejected before requests."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Mistral caches shared prefixes automatically and with_caching(key:) sends prompt_cache_key on Chat Completions. cached_tokens become cache_read. Conversations has neither the key nor cached token counts, so caching applies to Chat Completions only.","sources":["s823","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"parse_conversation_usage"}]},"cache_boundaries":{"v1":"na","v2":"unknown","offered":null,"notes":"The OpenAPI file gives Responses API content parts a prompt_cache_breakpoint field with mode explicit. The prompt caching guide and Responses guide describe only prompt_cache_key, and Chat Completions has no such field. A live request with a breakpoint partway through the prefix was accepted but changed nothing: automatic caching covered the same prefix with or without it. So no working behaviour is established either way. RubyLLM sends no boundary to Mistral.","sources":["s815","s823","s824"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"openai_prompt_caching?"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"mistral-small-latest","notes":"Responses requests with and without prompt_cache_breakpoint partway through a 29000 token prefix cached the same 29184 tokens."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"No cache resource endpoint appears in the OpenAPI inventory. Caching is automatic with an optional key.","sources":["s815","s823","s825"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"No inference endpoint exposes a compaction control. The Responses schema fixes truncation to disabled. Compaction inside Managed Agents belongs to the Agents SDK harness, a separate product.","sources":["s815","s824","s826"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"token_counting":{"v1":"na","v2":"native","offered":true,"notes":"chat.count_tokens renders the request as Responses input and posts it to POST /v1/responses/input_tokens, in public preview, with instructions, tools, structured output and images, whichever protocol the chat runs on. Conversations replies are sent as plain items, and thinking is counted as assistant text, since the endpoint takes no unsigned thinking. RubyLLM declares no other use of Mistral's Responses API.","sources":["s824","s827"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/input_tokens.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/token_counting.rb","method":"render_count_tokens_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb","method":"count_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"mistral-small-latest","spec":"spec/ruby_llm/providers/mistral_token_counting_spec.rb","example":"RubyLLM::Providers::Mistral token counting counts instructions, tools, structured output and images before generation","cassette":"spec/fixtures/vcr_cassettes/providers_mistral_token_counting_counts_instructions_tools_structured_output_and_images_before_generation.yml","notes":"Counts 22 tokens for the prompt, 156 with instructions, a tool and a schema, and 52 with an image."}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint calls the Conversations image_generation tool with a chat model and returns an Image with downloaded bytes and usage. One image per request; size and count raise.","sources":["s180","s301","s179"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/images.rb","method":"render_image_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/images.rb","method":"parse_image_responses"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations/images_spec.rb","example":"generates and downloads an image through paint","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_images_generates_and_downloads_an_image_through_paint.yml","notes":"RubyLLM.paint generated an image and downloaded its bytes."}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory and the image generation tool guide document generation from a prompt only. with: and mask: raise in RubyLLM.","sources":["s815","s180","s825"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/images.rb","method":"render_image_payload"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no video generation operation.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak posts to /v1/audio/speech, maps voice: to voice_id, and decodes the base64 audio. Saved voice IDs work as voice:. expressiveness, phoneme_hints and conversation_history go through provider_options.","sources":["s93","s828"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/speech.rb","method":"render_speech_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/speech.rb","method":"parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-mini-tts-latest","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality mistral/voxtral-mini-tts-latest can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_mistral_voxtral-mini-tts-latest_can_speak.yml","notes":"The recorded request sends voice_id."}]},"transcription":{"v1":"partial","v2":"native","offered":true,"notes":"RubyLLM.transcribe uploads the file to /v1/audio/transcriptions. prompt: maps to context_bias, timestamps: :segment to timestamp_granularities, and format: is rejected. Word timestamps are not offered.","sources":["s94","s829"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"render_transcription_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-mini-latest","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality mistral/voxtral-mini-latest can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_mistral_voxtral-mini-latest_can_transcribe_audio.yml","notes":"File upload and transcript text."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe with a block handles the SSE events transcription.text.delta, transcription.segment and transcription.done for an uploaded file, keeping text, speaker segments, duration and usage. Live audio over the realtime WebSocket endpoint has no integration.","sources":["s94","s829","s830"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"build_transcription_chunk"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"build_streamed_transcription"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-mini-latest","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality streams Mistral transcriptions with speaker segments and usage","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_streams_mistral_transcriptions_with_speaker_segments_and_usage.yml","notes":"The recording shows segment and done events."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"speaker_names: [] turns on diarize with segment timestamps, and segments carry speaker_id. Non-empty names and speaker_references: are rejected, although Mistral documents speaker_references for naming speakers from clips.","sources":["s94","s831"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-mini-latest","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality mistral/voxtral-mini-latest labels segments with speakers when speaker names are given","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_mistral_voxtral-mini-latest_labels_segments_with_speakers_when_speaker_names_are_given.yml","notes":"Recorded segments carry speaker_id."}]},"moderation":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.moderate posts to /v1/moderations and derives flagged from the returned category booleans. The /v1/chat/moderations endpoint for conversations is not used. No recording covers this route.","sources":["s832","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/moderation.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb","method":"render_moderation_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb","method":"parse_moderation_response"},{"version":"v2","path":"lib/ruby_llm/moderation.rb","method":"from_h"}]},"ocr":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.ocr returns typed pages and accepts pages:. Annotations, table_format, include_blocks and confidence scores go through provider_options. Uploaded file IDs cannot be passed as the document.","sources":["s95","s833"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/ocr.rb","method":"render_ocr_payload"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/ocr.rb","method":"parse_ocr_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-ocr-latest","spec":"spec/ruby_llm/ocr_spec.rb","example":"RubyLLM::OCR basic functionality mistral/mistral-ocr-latest extracts markdown from a PDF","cassette":"spec/fixtures/vcr_cassettes/ocr_basic_functionality_mistral_mistral-ocr-latest_extracts_markdown_from_a_pdf.yml","notes":"Local PDF processed through /v1/ocr."}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.embed posts to /v1/embeddings and dimensions: maps to output_dimension where the model allows it. output_dtype and encoding_format go through provider_options.","sources":["s185","s834"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/mistral/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/embeddings.rb","method":"render_embedding_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-embed","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality mistral/mistral-embed can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_mistral_mistral-embed_can_handle_a_single_text.yml","notes":"Single text embedding."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The embeddings endpoint accepts text as a string or array of strings. No image, audio or video input is documented.","sources":["s185","s815"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no rerank operation.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to /v1/files with purpose: and expires_in:, rounded up to whole hours as expiry. visibility goes through provider_options. uri: is rejected.","sources":["s96","s835"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/mistral/files.rb","method":"render_upload_fields"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/files.rb","method":"parse_file_response"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads mistral uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_mistral_uploads_a_pdf_through_the_files_api.yml","notes":"PDF uploaded with purpose ocr."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download fetches bytes from /files/:id/content, including files generated by hosted tools.","sources":["s96","s180"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"download_file_url"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/files.rb","method":"Files"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch submits inline requests to /v1/batch/jobs for /v1/chat/completions and reads results with inline=true. A job takes one model and cannot mix chat and embedding requests.","sources":["s835","s836"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions/batches.rb","method":"create_batch"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions/batches.rb","method":"batch_results"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with mistral/mistral-small-latest answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_mistral_mistral-small-latest_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Job submitted, polled to completion and results read inline."}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch sends embed_later requests to /v1/embeddings through the same job API and parses typed Embedding results. custom_id preserves scalar or array shape after reloading, and errors keep their slots. The recording covers submit, find and cancel; result parsing has unit specs.","sources":["s835","s836"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions/batches.rb","method":"mistral_batch_endpoint"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions/batches.rb","method":"parse_mistral_batch_body"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-embed","spec":"spec/ruby_llm/providers/mistral/chat_completions/batches_spec.rb","example":"RubyLLM::Providers::Mistral::ChatCompletions::Batches submits, reloads, and cancels an embedding batch","cassette":"spec/fixtures/vcr_cassettes/providers_mistral_chatcompletions_batches_submits_reloads_and_cancels_an_embedding_batch.yml","notes":"Submit, reload and cancel were recorded; the job did not complete."}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated advisor tool operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Perplexity Agent API, which is not an API of this provider.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"agent_skills":{"v1":"missing","v2":"missing","offered":true,"notes":"Mistral stores versioned skills through the beta /v2/skills API. RubyLLM registers no skill resource protocol, and no inference request takes a skill.","sources":["s176","s837"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated asynchronous research operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated background job operation. The Responses schema fixes background to false and responses are not stored. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s824","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated browser use operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"classification":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/classifications and /v1/chat/classifications run classifier models, and no RubyLLM operation calls them. RubyLLM.moderate covers only the moderation endpoint.","sources":["s838","s815"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb","method":"moderation_url"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated computer use operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated context editing operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated contextual embeddings operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"dedicated_transcription_api":{"v1":"partial","v2":"native","offered":true,"notes":"The dedicated /v1/audio/transcriptions endpoint is used through the Mistral transcription dialect. This row overlaps Transcription.","sources":["s94","s829"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated dubbing operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"fast_inference":{"v1":"na","v2":"passthrough","offered":true,"notes":"Priority Tier is selected with service_tier on completion requests, forwarded through with_provider_options. It is limited access: it needs an agreement and custom rate limits from Mistral. The served tier stays in the raw usage.","sources":["s839","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Mistral exposes file listing and deletion. RubyLLM Files implements upload, metadata lookup and download only.","sources":["s96"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/files.rb","method":"Files"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find"}]},"file_search_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Mistral Libraries provide managed indexing and retrieval. RubyLLM has no library creation or management API; its Files integration is separate.","sources":["s177","s360"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"fill_in_middle":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/fim/completions takes prompt and suffix for code models. RubyLLM has no operation for it, and raw options cannot change the chat endpoint.","sources":["s840","s815"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"fine_tuning":{"v1":"missing","v2":"na","offered":false,"notes":"Fine-tuning training is no longer available: GET /v1/fine_tuning/jobs returned 410 with \"Fine-tuning training has been deprecated and is no longer available. Existing fine-tuned models and LoRA adapter inference remain supported.\" RubyLLM has no fine-tuning lifecycle.","sources":["s99","s815","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/fine_tuning/jobs returned 410 Gone."}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google Maps grounding, which is not an API of this provider.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"guardrail_configuration":{"v1":"na","v2":"passthrough","offered":true,"notes":"Mistral documents a guardrails request field on Chat Completions and Conversations. with_provider_options forwards it on both, and results stay in the raw response. Agent-level guardrails belong to stored agents.","sources":["s832","s92","s841"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Mistral documents stored conversations through /v1/conversations with append, history and restart routes, and tool confirmations depend on them. RubyLLM sends store: false and keeps conversation state in the application.","sources":["s179","s362"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"store"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated image batch operation. The batch endpoint list names no image operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s835","s815","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names the Google Interactions API, which is not an API of this provider.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Mistral documents response_format json_object. with_provider_options forwards it and returns the text as usual; with_schema is the dedicated structured path.","sources":["s842","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/chat_request_options_spec.rb","example":"RubyLLM::Chat with params mistral/mistral-small-latest supports response_format param","cassette":"spec/fixtures/vcr_cassettes/chat_with_params_mistral_mistral-small-latest_supports_response_format_param.yml","notes":"The recorded request carries response_format json_object."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names Amazon Bedrock Mantle, which is not an API of this provider.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no compaction endpoint.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated music generation operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"nonchat_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"The batch endpoint list also covers FIM, moderation, OCR, classification, Conversations and transcription. RubyLLM batches only chat and embeddings.","sources":["s835","s836"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat_completions/batches.rb","method":"mistral_batch_endpoint"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"This row names a Google video endpoint, which is not an API of this provider.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated partner model protocol operation. Partner models such as Z.ai GLM are served through the same Mistral endpoints. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s841","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"prefix_completion":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Mistral documents prefix: true on a final assistant message. RubyLLM has no per-message option, so with_provider_options(messages: [...]) must replace the whole message list with raw values.","sources":["s810","s815"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/support/utils.rb","method":"deep_merge"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The Responses schema has allowed_callers, but no execution path exists for it: the endpoint rejects hosted tools, including code_interpreter, and a function with allowed_callers: [\"programmatic\"] was still called directly. The Responses guide and API description say only client-executed function tools are supported. RubyLLM has no Responses route for Mistral.","sources":["s815","s824","s827"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"mistral-small-latest","notes":"allowed_callers: [\"programmatic\"] was accepted and the function was called directly; a code_interpreter tool returned 400."}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching(key:) maps to prompt_cache_key on Chat Completions; other options raise. The Conversations route drops the setting because its schema has no key.","sources":["s823","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/chat.rb","method":"prompt_cache_options"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations/chat.rb","method":"render_payload"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated provider-managed fallback operation. Priority Tier falling back to Standard Tier is capacity routing, not model fallback. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s839","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Mistral documents realtime voice agents, realtime transcription and realtime speech over WebSocket sessions. RubyLLM has no Mistral realtime transport, and realtime conversations are outside 2.0 coverage.","sources":["s843","s830","s844","s845"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated response caching operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"responses":{"v1":"na","v2":"missing","offered":true,"notes":"POST /v1/responses is in public preview: stateless, with text, images, files, structured output, streaming, client-executed functions and reasoning. Mistral registers no :responses protocol in RubyLLM, so protocol: :responses is unavailable and raw options cannot reach the endpoint.","sources":["s824","s827","s841"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated Responses batch operation. The batch endpoint list excludes /v1/responses. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s835","s815","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated standalone web search operation. The Search Index search route queries an index you created and belongs with vector store management. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s846","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated apply patch tool operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"server_tool_image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:image_generation) works on default Chat Completions and on Conversations. Multi-completion and streamed responses keep every native message and generated attachment, without dispatching completed hosted calls as local tools.","sources":["s180","s301","s302"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/multi_completion.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/conversations.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/multi_completion.rb","method":"parse_multi_images"},{"version":"v2","path":"lib/ruby_llm/protocols/mistral/content.rb","method":"parse_generated_image"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/multi_completion_spec.rb","example":"streams and downloads a hosted image through the default chat API","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_multicompletion_streams_and_downloads_a_hosted_image_through_the_default_chat_api.yml","notes":"The hosted image tool streamed on Chat Completions and the image was downloaded."},{"status":"passed","recorded_on":"2026-10-11","model":"mistral-small-latest","spec":"spec/ruby_llm/protocols/mistral/conversations/images_spec.rb","example":"generates and downloads an image through paint","cassette":"spec/fixtures/vcr_cassettes/protocols_mistral_conversations_images_generates_and_downloads_an_image_through_paint.yml","notes":"RubyLLM.paint generated an image and downloaded its bytes."}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"POST /v1/responses rejects a tool_search tool with 400: this endpoint currently supports only client-executed function tools. defer_loading on a function is accepted and the function still runs. RubyLLM has no Responses route for Mistral.","sources":["s815","s824","s827"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"mistral-small-latest","notes":"tools with type tool_search returned 400: Tool types ['tool_search'] are not supported."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated X search tool operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated shell or computer tool operation. Sandboxes in Managed Agents belong to the Agents SDK harness. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s826","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated sound effects operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak with a block sends stream: true and decodes speech.audio.delta events into SpeechChunk values. speech.audio.done is required and its usage is kept. You send the complete text once; text streamed in over the realtime WebSocket endpoint has no integration.","sources":["s847","s93","s844"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/speech.rb","method":"stream_speech"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voxtral-mini-tts-latest","spec":"spec/ruby_llm/speech_streaming_spec.rb","example":"RubyLLM::Speech streams speech from mistral","cassette":"spec/fixtures/vcr_cassettes/speech_streams_speech_from_mistral.yml","notes":"Streamed audio deltas and final usage."}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated task budget operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated text analysis operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"vector_store_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Mistral Libraries, the RAG search-index API and the Search Index API support indexing and retrieval. RubyLLM can pass an existing library to a hosted tool, but has no index creation or management API.","sources":["s177","s178","s848","s846"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated video batch operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated video editing operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated video extension operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"voice_cloning":{"v1":"missing","v2":"passthrough","offered":true,"notes":"ref_audio clones a voice for one request and RubyLLM.speak forwards it through provider_options. Mistral also has saved voices at /v1/audio/voices; a saved voice ID works as voice:, but creating and managing voices has no RubyLLM lifecycle.","sources":["s849","s93","s181"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral/speech.rb","method":"render_speech_payload"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated voice conversion operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated standalone web fetch operation. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s815","s825","s92"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The API inventory documents no dedicated Responses over WebSocket operation. The Responses API uses HTTP with server-sent events. WebSocket sessions cover realtime audio, recorded under Realtime sessions. Scope: Mistral inference, Conversations, connector, library and skill APIs. Vibe, the Agents SDK harness and client-side Search Toolkit helpers are separate.","sources":["s824","s815","s830"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/mistral.rb"},{"version":"v2","path":"lib/ruby_llm/providers/mistral.rb"}]}},"gaps":[{"notes":"Agent skills: Mistral stores versioned skills through the beta /v2/skills API. RubyLLM registers no skill resource protocol.","sources":["s176"]},{"notes":"Citations: reference chunks with reference_ids are mapped on Chat Completions only; Conversations reference chunks are not checked, and streamed Chat Completions citations carry no offset.","sources":["s817"]},{"notes":"Classification: The classification endpoints have no RubyLLM operation.","sources":["s838"]},{"notes":"File listing and deletion: Mistral exposes both, while RubyLLM Files implements upload, metadata lookup and download only.","sources":["s96"]},{"notes":"File search store management, Vector store management: Mistral Libraries and Search Index APIs support indexing and retrieval. RubyLLM can pass an existing library into a hosted tool, but has no creation or management API.","sources":["s177","s178","s848"]},{"notes":"Fill-in-the-middle completion: The /v1/fim/completions endpoint has no RubyLLM operation.","sources":["s840"]},{"notes":"Other batch operations: Mistral also batches FIM, moderation, OCR, classification, Conversations and transcription.","sources":["s835"]},{"notes":"Responses API: The public preview /v1/responses endpoint has no registered RubyLLM protocol. Token counting uses /v1/responses/input_tokens directly.","sources":["s824"]},{"notes":"Fast inference, Guardrail configuration, JSON object mode, Prefix completion: Mistral documents service_tier, guardrails, json_object and assistant prefix. They work only through with_provider_options; prefix must replace the whole message list. Priority Tier is limited access.","sources":["s839","s832","s842","s810"]},{"notes":"Voice cloning: ref_audio works through provider_options. Creating and managing saved voices has no RubyLLM lifecycle.","sources":["s849","s181"]},{"notes":"PDF input, OCR / document parsing: Uploaded file IDs cannot be attached; use a URL or local file.","sources":["s95","s815"]},{"notes":"Speaker identification: speaker_references: and non-empty speaker_names: are rejected, although Mistral documents reference clips for naming speakers.","sources":["s831"]},{"notes":"Streaming transcription, Streaming speech generation: SSE streaming works. The realtime WebSocket endpoints for live audio and text input have no integration.","sources":["s830","s844"]},{"notes":"Prompt caching, Prompt cache options: These apply to Chat Completions. On Conversations, with_caching has no effect and cached tokens are not reported.","sources":["s823","s179"]},{"notes":"Web search tool, Code execution tool, File search tool, Remote MCP tools: Mistral says the Agents SDK replaces the Agents and Conversations API and calls Agents API v1 about to be deprecated, with no date. RubyLLM uses Conversations for the hosted tools that Chat Completions does not offer.","sources":["s850","s819"]},{"notes":"Cache boundaries: the Responses schema has a breakpoint field, and a live request showed no difference from automatic caching.","sources":["s815","s824"]}]},{"id":"openrouter","name":"OpenRouter","notes":"The audit covers Chat Completions, Responses, Messages, Decisions, Extractions, media, Files, Batch, routing and documented server tools, for the models OpenRouter routes from other providers. RubyLLM defaults to Chat Completions and selects Responses on request; it has no Messages route. Workspace administration is sampled, not exhaustively audited.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"Default route is Chat Completions; protocol: :responses selects Responses. Applies to the models OpenRouter routes.","sources":["s851","s276"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality openrouter/claude-haiku-4-5 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_openrouter_claude-haiku-4-5_can_have_a_basic_conversation.yml","notes":"Completion over /chat/completions."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Server-sent events on Chat Completions and Responses, with reasoning details, usage and reported cost on chunks.","sources":["s852","s851"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/streaming.rb","method":"build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses openrouter/claude-haiku-4-5 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_openrouter_claude-haiku-4-5_supports_streaming_responses.yml","notes":"Stream over /chat/completions."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends response_format json_schema on supporting models. RubyLLM removes the strict key from the schema body.","sources":["s853","s854"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"strip_schema_strict"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat with_schema with openrouter/claude-haiku-4-5 accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_openrouter_claude-haiku-4-5_accepts_a_json_schema_and_returns_structured_output.yml","notes":"JSON schema request and parsed output."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as image_url parts. Local files and remote URLs both work. Availability depends on the upstream model.","sources":["s855"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/media.rb","method":"render_content"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models openrouter/claude-haiku-4-5 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_openrouter_claude-haiku-4-5_can_understand_local_images.yml","notes":"Local image over /chat/completions."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"PDFs render as file parts on Chat Completions and input_file on Responses. A PDF above the inline threshold can upload as a provider file. Upstream support or the file-parser plugin decides how it is read.","sources":["s275"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/media.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"provider_file_attachable?"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"google/gemini-2.5-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models openrouter/gemini-2.5-flash understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_openrouter_gemini-2_5-flash_understands_pdfs.yml","notes":"Local PDF over /chat/completions."}]},"audio_input":{"v1":"native","v2":"native","offered":true,"notes":"Audio attachments render as base64 input_audio parts. The audio guide shows Responses and the Chat Completions schema carries the same part. Audio URLs are not accepted. No OpenRouter recording exercises this path.","sources":["s118","s856"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/media.rb","method":"render_content"}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"Video attachments render as video_url parts, from a URL or a base64 data URL, on supporting models. Which URLs work depends on the upstream provider. Generation is a separate operation. No OpenRouter recording exercises this path.","sources":["s857"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/media.rb","method":"render_video"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Reasoning text, summaries, signatures and encrypted details are parsed, and reasoning_details replay unchanged across tool calls.","sources":["s858"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"parse_thinking_text, parse_raw_reasoning"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking openrouter/claude-haiku-4-5 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_openrouter_claude-haiku-4-5_returns_thinking_when_available.yml","notes":"Reasoning returned on /chat/completions."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking maps effort to reasoning.effort, budget to reasoning.max_tokens and the switch to reasoning.enabled.","sources":["s858"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"build_reasoning"}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"URL annotations and root citations normalize to Citation objects. Web search through the server tool or a search model produces them.","sources":["s859"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"extract_citations"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"perplexity/sonar","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with openrouter/perplexity/sonar returns search result citations","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_openrouter_perplexity_sonar_returns_search_result_citations.yml","notes":"Search result citations from a search model."}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools on Chat Completions and Responses, with streaming and multi-turn tool loops. Availability depends on the upstream model.","sources":["s860"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling openrouter/claude-haiku-4-5 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_openrouter_claude-haiku-4-5_can_use_tools.yml","notes":"Tool call and result over /chat/completions."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) maps to tool_choice. For choice: :none RubyLLM removes tools from the payload.","sources":["s860","s854"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control openrouter/claude-haiku-4-5 respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_openrouter_claude-haiku-4-5_respects_specific_tool_choice.yml","notes":"Forced tool call."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Several tool calls in one response are parsed and run. Whether a model emits several is up to the model.","sources":["s860"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"anthropic/claude-haiku-4.5","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling openrouter/claude-haiku-4-5 can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_openrouter_claude-haiku-4-5_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Several tool calls in one response."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) sends parallel_tool_calls false. Upstream support varies by model.","sources":["s860","s854"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"upstage/solar-pro4","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control openrouter/upstage/solar-pro4 respects calls one for sequential execution","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_openrouter_upstage_solar-pro4_respects_calls_one_for_sequential_execution.yml","notes":"Sequential tool calls with parallel_tool_calls false."}]},"server_web_search":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends openrouter:web_search on Chat Completions and Responses. Results surface as citations and usage counters. Server tools are beta.","sources":["s269","s859"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-5.2","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat with openrouter/openai/gpt-5.2 searches transparently, returning citations and usage counters","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_openrouter_openai_gpt-5_2_searches_transparently_returning_citations_and_usage_counters.yml","notes":"Citations and web_search_requests in usage."}]},"server_web_fetch":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:web_fetch) or :url_context sends openrouter:web_fetch, and options such as allowed_domains pass through the alias. No OpenRouter recording exercises it. Server tools are beta.","sources":["s269","s861"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"SERVER_TOOL_ALIASES"}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:code_execution) on protocol: :responses runs the hosted shell, and output, exit status and billing are preserved. The docs say this tool returns 400 on Chat Completions, the default route, so select Responses. Server tools are beta.","sources":["s269","s270"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"server_tool_aliases"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-5.2","spec":"spec/ruby_llm/protocols/openrouter/responses_spec.rb","example":"RubyLLM::Protocols::OpenRouter::Responses executes a hosted shell and returns its real output and billed usage","cassette":"spec/fixtures/vcr_cassettes/protocols_openrouter_responses_executes_a_hosted_shell_and_returns_its_real_output_and_billed_usage.yml","notes":"Runs python3 in the hosted shell through openrouter:shell. stdout 391, exit code 0 and billed usage are in the recording."}]},"server_file_search":{"v1":"unknown","v2":"unknown","offered":null,"notes":"The Responses schema lists a file_search tool with vector_store_ids, but no guide describes it, GET /tools omits it, and the inventory has no vector store endpoint. A live request without a provider key returned 400: file_search vector_store_ids are only supported for BYOK requests. Whether it runs with a provider key, and for which models, stays unverified, because no OpenRouter BYOK setup was available.","sources":["s273","s269","s274"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-4.1-nano","notes":"POST /responses with file_search returned 400: code_interpreter container ids, container file_ids and file_search vector_store_ids are only supported for BYOK requests."}]},"server_mcp":{"v1":"unknown","v2":"native","offered":true,"notes":"with_provider_tools(mcp:) on protocol: :responses invokes a remote server. Full and streamed responses carry tool names, arguments and results as ServerToolCall values, and approval works: require_approval: always pauses for chat.pending_questions, and approving or denying continues from the local transcript. The guides do not describe hosted MCP; the request schema and these recordings do. OpenRouter's stream reports only arguments until the call finishes, so its first mcp_call chunk has no name or result.","sources":["s273","s269"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"SERVER_TOOL_ALIASES, build_chunk"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/approvals.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.2","spec":"spec/ruby_llm/protocols/openrouter/responses_spec.rb","example":"RubyLLM::Protocols::OpenRouter::Responses records the remote MCP calls it runs, with their tool names and results","cassette":"spec/fixtures/vcr_cassettes/protocols_openrouter_responses_records_the_remote_mcp_calls_it_runs_with_their_tool_names_and_results.yml","notes":"Streamed mcp_call with name, arguments and result on the final message."},{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.2","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openrouter executes a remote read-only tool after approval using stateless continuation","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openrouter_executes_a_remote_read-only_tool_after_approval_using_stateless_continuation.yml","notes":"An approval pause, then continuation built from local history."},{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.2","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openrouter denies a remote call without executing it","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openrouter_denies_a_remote_call_without_executing_it.yml","notes":"A denied call finishes without running the tool."},{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.2","spec":"spec/ruby_llm/chat_mcp_spec.rb","example":"RubyLLM::Chat remote MCP tools openrouter streams a remote approval and its completed result","cassette":"spec/fixtures/vcr_cassettes/chat_remote_mcp_tools_openrouter_streams_a_remote_approval_and_its_completed_result.yml","notes":"Streamed approval pause and completed result."}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Most upstream providers cache automatically, and cache read and write counts are parsed from usage. with_caching adds a top-level cache_control for the models that need it.","sources":["s122"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"cache_read_tokens"},{"version":"v1","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"render_payload"}]},"cache_boundaries":{"v1":"passthrough","v2":"native","offered":true,"notes":"cache_until_here adds a cache_control block on supporting upstream models. OpenAI explicit caching uses prompt_cache_breakpoint, which this marker does not send; provider_options cannot place it on a content block.","sources":["s122"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"inject_cache_control"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_cache_lifetimes_spec.rb","example":"sends OpenRouter the spelling its cache_control takes","notes":"Renders cache_control with the requested lifetime."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. Prompt caching is automatic or marked per request. No separately created cache resource exists.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"compaction":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_compaction adds the context-compression plugin, which removes or truncates middle messages when the prompt exceeds the context window. It does not summarize and has no threshold, so portable options are dropped with a debug log.","sources":["s123"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"apply_compaction"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_compaction_spec.rb","example":"maps to the context-compression plugin on OpenRouter","notes":"Adds the context-compression plugin to the payload."}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. No token counting endpoint appears in the OpenAPI paths. Counts arrive in response usage.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"image_generation":{"v1":"native","v2":"native","offered":true,"notes":"paint uses /images and reads the reported cost. The endpoint accepts n up to 10, but RubyLLM rejects count: and reads the first image only. Streaming partial images is not exposed.","sources":["s126"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter/images.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/images.rb","method":"render_image_payload, parse_image_responses"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"google/gemini-3.1-flash-lite-image","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality openrouter/google/gemini-3.1-flash-lite-image can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_openrouter_google_gemini-3_1-flash-lite-image_can_paint_images.yml","notes":"Image returned by /images."}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"paint(with:) sends reference images as input_references. mask: raises because the API has no mask field; markup, its overlay for some models, needs provider_options.","sources":["s126"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/images.rb","method":"build_input_references, validate_paint_inputs!"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"google/gemini-3.1-flash-lite-image","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image edit functionality with openrouter reference images openrouter/google/gemini-3.1-flash-lite-image edits images passed via with","cassette":"spec/fixtures/vcr_cassettes/image_edit_functionality_with_openrouter_reference_images_openrouter_google_gemini-3_1-flash-lite-image_edits_images_passed_via_with.yml","notes":"Reference image edit through /images."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate submits to /videos, polls the job and downloads the content. Up to two with: images become first and last frames. Only unit specs cover this path.","sources":["s119","s281"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/videos.rb","method":"render_video_payload, download_video"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/openrouter/videos_spec.rb","example":"maps reference images to first and last frames","notes":"Covers payload, job status, cost and download with stubbed responses."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak posts to /audio/speech and returns typed audio. Voices are model-specific, so no default voice is assumed.","sources":["s127"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/speech.rb","method":"render_speech_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"hexgrad/kokoro-82m","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality openrouter/hexgrad/kokoro-82m can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_openrouter_hexgrad_kokoro-82m_can_speak.yml","notes":"Audio bytes from /audio/speech."}]},"transcription":{"v1":"partial","v2":"native","offered":true,"notes":"transcribe posts base64 audio to /audio/transcriptions. Formats are json and verbose_json. prompt: raises because the API ignores it.","sources":["s128"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/transcription.rb","method":"transcribe"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-4o-mini-transcribe","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openrouter/openai/gpt-4o-mini-transcribe can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openrouter_openai_gpt-4o-mini-transcribe_can_transcribe_audio.yml","notes":"Transcribes the recorded fixture through /audio/transcriptions."}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The transcription reference documents one completed JSON response, with no stream flag or delta events. RubyLLM raises when given a block.","sources":["s128","s862"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/transcription.rb","method":"raise_transcription_streaming_unsupported"}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"speaker_names: [] sets the Azure and Deepgram diarization options under provider.options and asks for verbose_json. Numbered speaker labels are preserved. Named speakers and reference audio raise because the API has no field for them. Backend support varies.","sources":["s128"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/transcription.rb","method":"transcribe, validate_transcription_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"microsoft/mai-transcribe-2","spec":"spec/ruby_llm/protocols/openrouter/transcription_spec.rb","example":"RubyLLM::Protocols::OpenRouter::Transcription transcribes a real recording with speaker labels and reported duration and cost","cassette":"spec/fixtures/vcr_cassettes/protocols_openrouter_transcription_transcribes_a_real_recording_with_speaker_labels_and_reported_duration_and_cost.yml","notes":"Speaker labels, duration and reported cost in a verbose_json transcript."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The OpenAPI paths have no moderation endpoint. Guardrail policies, Armor and Alignment apply rules to requests and tool calls, not to arbitrary content on demand. Those configuration APIs are under Guardrail configuration.","sources":["s274","s856","s863"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"OpenRouter has no route that returns the text of a page. POST /api/alpha/extractions is the closest: it answers a fields map with typed values on a few extraction models, which is a different operation from OCR, and it is alpha. OCR also runs inside the chat file-parser plugin, which counts as PDF input.","sources":["s864","s275","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"embed posts to /embeddings and returns vectors, usage and the reported cost.","sources":["s304","s865"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/embeddings.rb","method":"render_embedding_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/text-embedding-3-small","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding basic functionality openrouter/openai/text-embedding-3-small can handle a single text","cassette":"spec/fixtures/vcr_cassettes/embedding_basic_functionality_openrouter_openai_text-embedding-3-small_can_handle_a_single_text.yml","notes":"Single text embedding."}]},"multimodal_embeddings":{"v1":"passthrough","v2":"native","offered":true,"notes":"embed(with:) renders one content array with image_url, input_audio, input_video and input_file parts and returns one vector. Supported media depend on the model. Text arrays cannot be combined with attachments.","sources":["s304","s865"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/embeddings.rb","method":"render_embedding_content"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"google/gemini-embedding-2","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings openrouter/google/gemini-embedding-2 embeds an image alongside text","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_openrouter_google_gemini-embedding-2_embeds_an_image_alongside_text.yml","notes":"Text and image joined in one input."},{"status":"passed","recorded_on":"2026-10-11","model":"google/gemini-embedding-2","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings openrouter/google/gemini-embedding-2 embeds ruby.mp4","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_openrouter_google_gemini-embedding-2_embeds_ruby_mp4.yml","notes":"Video input."}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"rerank posts to /rerank and returns ordered results with the reported cost.","sources":["s866"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"Rerank"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"voyageai/rerank-2.5-lite","spec":"spec/ruby_llm/rerank_spec.rb","example":"RubyLLM::Rerank reranking with openrouter/voyageai/rerank-2.5-lite orders documents by relevance and reports the exact cost","cassette":"spec/fixtures/vcr_cassettes/rerank_reranking_with_openrouter_voyageai_rerank-2_5-lite_orders_documents_by_relevance_and_reports_the_exact_cost.yml","notes":"Ordered documents and exact cost."}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"upload posts to the beta /files endpoint and find reads metadata. Files can be up to 100 MiB and belong to a workspace.","sources":["s279"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/files.rb","method":"parse_file_response"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/files_spec.rb","example":"normalizes file metadata","notes":"Parses the file response with stubbed HTTP."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"download reads /files/{id}/content. OpenRouter only serves files a model or sandbox created (downloadable: true); a file you uploaded returns 400. Container file endpoints are not integrated.","sources":["s279"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"download"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/files.rb","method":"parse_file_response"}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM submits inline requests with one model and one endpoint to /api/v1/batches, then finds the batch and reads ordered results, per-request failures and the aggregate cost. Public image URLs go through as URLs; inline media is left to OpenRouter, which accepts it only on some accounts and providers. Models need a batch endpoint, and some of those need your own provider key. Turnaround depends on the provider: Google finished in minutes, Anthropic had not finished after 30 minutes. callback_url and provider.only are not exposed. When a whole batch fails, RubyLLM reports the failed status without OpenRouter's reason. OpenRouter has no cancel operation.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"batch_api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/batches.rb","method":"create_batch, batch_results"},{"version":"v2","path":"lib/ruby_llm/batch.rb","method":"cost"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"google/gemini-3.7-flash","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openrouter/google/gemini-3.7-flash answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_openrouter_google_gemini-3_7-flash_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Two staged chats completed and were appended to their conversations."},{"status":"passed","recorded_on":"2026-11-10","model":"google/gemini-3.7-flash","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openrouter/google/gemini-3.7-flash attachments reads a public image URL","cassette":"spec/fixtures/vcr_cassettes/batch_with_openrouter_google_gemini-3_7-flash_attachments_reads_a_public_image_url.yml","notes":"A public image URL was read through a batch."},{"status":"unit","spec":"spec/ruby_llm/protocols/openrouter/batches_spec.rb","notes":"Request and result contract, ordering, per-request failures and aggregate cost."}]},"embedding_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM submits inline requests with one model and one endpoint to /api/v1/batches, then finds the batch and reads ordered results, per-request failures and the aggregate cost. Text input only, without input_type or provider, matching the documented limits. Some models need your own provider key, and on Google each request takes one input.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"batch_api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/batches.rb","method":"create_batch, batch_results"},{"version":"v2","path":"lib/ruby_llm/batch.rb","method":"cost"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"google/gemini-embedding-001","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with openrouter/google/gemini-embedding-001 embeddings embeds staged texts and hydrates each request result","cassette":"spec/fixtures/vcr_cassettes/batch_with_openrouter_google_gemini-embedding-001_embeddings_embeds_staged_texts_and_hydrates_each_request_result.yml","notes":"Two staged embeddings completed, with vectors and the reported cost."},{"status":"unit","spec":"spec/ruby_llm/protocols/openrouter/batches_spec.rb","notes":"Request and result contract, ordering, per-request failures and aggregate cost."}]},"advisor_tool":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"Pass a raw { type: \"openrouter:advisor\", parameters: { model: ... } } hash to with_provider_tools. RubyLLM forwards it and has no alias or typed advisor result. Available on Chat Completions, Responses and Messages.","sources":["s868","s269"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"add_raw"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"SERVER_TOOL_ALIASES"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names Perplexity's Agent API. It is not a capability of this provider.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"unknown","offered":null,"notes":"The Messages response schema carries a container object with a skills list, and the request schema has no such field. A plain request with container.skills is accepted and returns container: null, and adding Anthropic's code_execution tool, which skills need, returned 400 Unknown server-tool shorthand, so no request could actually run a skill. No guide documents uploading or attaching skills. The Agent SDK skills example runs on the client.","sources":["s869","s274"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"anthropic/claude-opus-4.8","notes":"A plain Messages request with container.skills returned container: null. Adding a code_execution_20250825 tool returned 400: Unknown server-tool shorthand."}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. Long-running research examples in the Agent SDK run on the client. /videos and /batches are the only asynchronous job APIs.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"background_responses":{"v1":"unknown","v2":"na","offered":false,"notes":"background: true is accepted and ignored: the response comes back completed in the same request, with background: false, and GET /responses/{id} returned 404. POST /responses is the only documented Responses route.","sources":["s273","s276"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-4.1-nano","notes":"POST /responses with background: true returned status completed and background: false. Retrieving the id returned 404."}]},"browser_use":{"v1":"unknown","v2":"na","offered":false,"notes":"No browser tool exists on this API: GET /tools lists none, and a Responses request with openrouter:browser_use was rejected the same way as a nonsense tool. The Responses schema defines the output item only.","sources":["s273","s269"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-4.1-nano","notes":"GET /tools lists no browser tool. openrouter:browser_use and openrouter:nonsense both returned 400 from the upstream."}]},"classification":{"v1":"missing","v2":"native","offered":true,"notes":"judge sends choice, probability and score questions to /api/alpha/decisions and returns typed answers with probabilities on decision models such as Jev. The route is alpha. Custom Classifiers, the separate workspace feature that tags logged generations, are configured outside RubyLLM.","sources":["s870","s871","s277"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/judgments.rb","method":"judgment_url"},{"version":"v2","path":"lib/ruby_llm/protocols/system_one/judgments.rb","method":"TYPES"}],"verification":[{"status":"passed","recorded_on":"2026-10-21","model":"typesafe/jev-1.14","spec":"spec/ruby_llm/providers/openrouter/judgments_spec.rb","example":"RubyLLM::Providers::OpenRouter::Judgments with the Decisions API judges all three question types through the compact DSL","cassette":"spec/fixtures/vcr_cassettes/providers_openrouter_judgments_with_the_decisions_api_judges_all_three_question_types_through_the_compact_dsl.yml","notes":"A choice question returns one label with a probability for each."}]},"computer_use":{"v1":"unknown","v2":"unknown","offered":null,"notes":"The Responses schema lists a computer_use_preview tool, but GET /tools omits it and no guide establishes a working upstream integration. Live requests to OpenAI models through OpenRouter returned 400 for computer_use_preview on gpt-4.1-nano, gpt-5.4-mini and gpt-5.6-luna, and openai/computer-use-preview is not a valid model id. A computer tool on gpt-5.6-luna returned reasoning only, with no computer_call. Whether it runs stays unverified.","sources":["s273","s269"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-5.6-luna","notes":"computer_use_preview returned 400 upstream; a computer tool returned reasoning output and no computer_call."}]},"context_editing":{"v1":"na","v2":"missing","offered":true,"notes":"The Messages API applies context_management, in Anthropic's shape: a clear_tool_uses_20250919 edit returned applied_edits and cut input tokens from 10587 to 2612 on anthropic/claude-haiku-4.5. RubyLLM has no Messages route on this provider.","sources":["s869","s274"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"anthropic/claude-haiku-4.5","notes":"POST /messages with context_management, on both default and Anthropic-only routing, returned applied_edits with 8100 input tokens and 5 tool uses cleared."}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. The embeddings endpoint takes input_type and has no document-context field.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"partial","v2":"native","offered":true,"notes":"POST /audio/transcriptions is a dedicated JSON and multipart endpoint. RubyLLM uses the JSON form.","sources":["s128"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/transcription.rb","method":"transcribe"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-4o-mini-transcribe","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality openrouter/openai/gpt-4o-mini-transcribe can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_openrouter_openai_gpt-4o-mini-transcribe_can_transcribe_audio.yml","notes":"Transcribes the recorded fixture through /audio/transcriptions."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. ","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"fast_inference":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"service_tier priority (alias fast) on a request is passed through with_provider_options, and RubyLLM has no speed API. The :nitro model variant also favors throughput. Support and billing vary by model.","sources":["s278","s273"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"GET /files and DELETE /files/{id} exist. RubyLLM implements upload, metadata lookup and download only.","sources":["s279","s872"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/files.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. The OpenAPI paths have no vector or file-search store route.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. The OpenAPI paths list no /completions route and no suffix parameter.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"OpenRouter routes to public and private endpoints and has no training job API. Tuning at an upstream provider is outside this inventory.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google Maps grounding. The inventory and the web-search guide do not mention it.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"missing","v2":"missing","offered":true,"notes":"/guardrails and its assignment routes create and assign policies for budgets, model access and content moderation. RubyLLM does not manage them. Request routing preferences do not replace this administrative API.","sources":["s863","s873"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The Responses guide says requests are stateless: store: true and a non-null previous_response_id return 400. Agent SDK conversation helpers run on the client.","sources":["s276","s274"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Batch API returns text and embeddings only and rejects image output. Image URLs are accepted as input to chat requests, which is the chat batches row.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names Google's Interactions API. It is not a capability of this provider.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"response_format { type: \"json_object\" } goes through with_provider_options on supporting models. with_schema is the separate JSON schema operation.","sources":["s854"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names Bedrock Mantle protocols. It is not a capability of this provider.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"missing","offered":true,"notes":"A compaction_trigger input item returned a compaction output item from openai/gpt-5.4-mini with max_output_tokens of 20000, the minimum the upstream demands. A request to anthropic/claude-haiku-4.5 at 60 tokens returned no output, which does not test it. No guide documents it and there is no compact route. RubyLLM has no integration for this item here.","sources":["s273","s274"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"openai/gpt-5.4-mini","notes":"POST /responses with a compaction_trigger item and max_output_tokens: 20000 returned a compaction item."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. Seed Audio prompts describe speech, sound effects, ambience or a scene; the guide does not cover music.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Batches cover Chat Completions, Responses, Messages and embeddings. No other operation can be batched.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"omni_video_generation":{"v1":"missing","v2":"passthrough","offered":true,"notes":"input_references takes image, audio and video assets; audio and video are honored by providers such as BytePlus Seedance generation 2 and newer. animate maps only first and last frames, so references go through provider_options.","sources":["s281","s119"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/videos.rb","method":"render_video_payload"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/openrouter/videos_spec.rb","example":"sends model, prompt, and provider options","notes":"provider_options reach the video payload."}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. /messages is an Anthropic-shaped route over any model, not a publisher-specific protocol for partner models.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"prefix_completion":{"v1":"unknown","v2":"passthrough","offered":true,"notes":"A final assistant message is continued on models that allow it; newer Claude models reject it. Chat#complete treats a trailing assistant message as answered and sends nothing, so the messages array has to go through with_provider_options.","sources":["s851","s874"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"answered?, with_provider_options"}]},"programmatic_tool_calling":{"v1":"na","v2":"unknown","offered":null,"notes":"Responses function tools accept allowed_callers with a programmatic value, in the schema only. The one guide that names the field, Advisor, says it is ignored there. Live, gpt-5.4-mini rejected a programmatic_tool_calling tool, and gpt-5.6-luna accepted it and still called the function directly, with no program caller. Whether any model honors it stays unverified.","sources":["s273","s868"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-5.6-luna","notes":"A function with allowed_callers: [\"programmatic\"] and a programmatic_tool_calling tool returned a function call with no caller field."}]},"prompt_cache_controls":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_caching maps ttl to cache_control. Other options raise; prompt_cache_key and prompt_cache_options go through with_provider_options.","sources":["s122"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/providers/openrouter/chat.rb","method":"prompt_cache_options, prompt_cache_control"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/chat_cache_lifetimes_spec.rb","example":"sends OpenRouter the spelling its cache_control takes","notes":"Renders cache_control with the requested lifetime."}]},"provider_server_fallback":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"The models list and provider preferences go through with_provider_options. Chat#with_fallbacks is RubyLLM's own client-side feature.","sources":["s282"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents HTTP and SSE only. Agent SDK streaming and microphone examples run on the client.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"response_caching":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"The X-OpenRouter-Cache header goes through with_headers. This whole-response cache is separate from prompt caching.","sources":["s130"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_headers"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"protocol: :responses selects the OpenRouter Responses dialect, while ordinary chats stay on Chat Completions. Hosted output, reasoning, citations, reported cost and usage are normalized. Judgments keep their Decisions route.","sources":["s276","s273"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol, resolve_protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"parse_usage"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-5.2","spec":"spec/ruby_llm/protocols/openrouter/responses_spec.rb","example":"RubyLLM::Protocols::OpenRouter::Responses executes a hosted shell and returns its real output and billed usage","cassette":"spec/fixtures/vcr_cassettes/protocols_openrouter_responses_executes_a_hosted_shell_and_returns_its_real_output_and_billed_usage.yml","notes":"Runs python3 in the hosted shell through openrouter:shell. stdout 391, exit code 0 and billed usage are in the recording."}]},"responses_batches":{"v1":"missing","v2":"partial","offered":true,"notes":"RubyLLM submits inline requests with one model and one endpoint, then finds the batch and reads ordered results, per-request failures and the aggregate cost. It posts to /api/beta/batches, and the docs document /api/v1/batches. No batch has completed through a live recording. A chat on protocol: :responses renders /v1/responses rows, and unit specs cover reading those results. Media parts raise, and callback_url and provider.only are not exposed.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol, batch_api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/batches.rb","method":"batch_request_endpoint, batch_parser"}],"verification":[{"model":"anthropic/claude-haiku-4.5, openai/gpt-4.1-nano:batch, openai/text-embedding-3-small","spec":"spec/ruby_llm/protocols/openrouter/batches_spec.rb","status":"unavailable","checked_on":"2026-09-07","notes":"Fresh actual API probes saved as observed model/status/error tuples, not a passing VCR batch test No batch ID returned, no resources created; Catalog and RubyLLM registry confirmed the explicit :batch variant before its one submission; Stopped submissions after the catalog contradiction; no brute-force model guessing Observed provider limitation: [{\"model\": \"anthropic/claude-haiku-4.5\", \"endpoint\": \"/v1/chat/completions\", \"status\": 400, \"error\": \"Model 'anthropic/claude-haiku-4.5' does not have a :batch endpoint.\"}, {\"model\": \"openai/text-embedding-3-small\", \"endpoint\": \"/v1/embeddings\", \"status\": 400, \"error\": \"Model 'openai/text-embedding-3-small' does not have a :batch endpoint.\"}, {\"model\": \"openai/gpt-4.1-nano:batch\", \"endpoint\": \"/v1/chat/completions\", \"status\": 400, \"error\": \"Model 'openai/gpt-4.1-nano:batch' does not have a :batch endpoint.\"}]"},{"status":"unit","spec":"spec/ruby_llm/protocols/openrouter/batches_spec.rb","notes":"Documented request/result contract and protocol regression coverage; no completed live batch or patch execution claimed."}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. Search exists only as a server tool during generation.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"server_apply_patch":{"v1":"missing","v2":"partial","offered":true,"notes":"with_provider_tools(:apply_patch) on protocol: :responses declares the tool, and output items are kept. OpenRouter validates the diff but never applies it: the application must return apply_patch_call_output, and RubyLLM has no patch execution or result lifecycle. The docs say Chat Completions does not support this tool.","sources":["s269","s380"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"server_tool_aliases"}]},"server_tool_image_generation":{"v1":"passthrough","v2":"native","offered":true,"notes":"with_provider_tools(:image_generation) sends openrouter:image_generation, with options passed through the alias. The image URL goes to the model, so results reach you through its message. No OpenRouter recording exercises it.","sources":["s269","s875"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"SERVER_TOOL_ALIASES"}]},"server_tool_search":{"v1":"missing","v2":"partial","offered":true,"notes":"with_tools(defer: true) on protocol: :responses sends defer_loading with a bare tool_search tool. The docs and request schema only list openrouter:tool_search, RubyLLM has no alias for it, and no recording exercises this route. On Chat Completions, RubyLLM runs its own client-side tool_search instead of the server tool.","sources":["s272","s269","s273"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"NATIVE_TOOL_SEARCH, apply_tool_search"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/tools/search.rb"}]},"server_x_search":{"v1":"na","v2":"native","offered":true,"notes":"with_provider_tools(web_search: { parameters: { x_search: {} } }) passes the documented opt-in through the alias. It applies only to native search on SpaceXAI models, and X results bill per item. No recording exercises it.","sources":["s859"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"merge_tool_options"}]},"shell_computer_tools":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:shell) or :code_execution on protocol: :responses runs the hosted shell, and RubyLLM preserves action, output and exit status. openrouter:bash is Messages only, which RubyLLM does not route to, and no interactive desktop tool is documented.","sources":["s269","s270","s271"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openrouter/responses.rb","method":"server_tool_aliases"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"openai/gpt-5.2","spec":"spec/ruby_llm/protocols/openrouter/responses_spec.rb","example":"RubyLLM::Protocols::OpenRouter::Responses executes a hosted shell and returns its real output and billed usage","cassette":"spec/fixtures/vcr_cassettes/protocols_openrouter_responses_executes_a_hosted_shell_and_returns_its_real_output_and_billed_usage.yml","notes":"Runs python3 in the hosted shell through openrouter:shell. stdout 391, exit code 0 and billed usage are in the recording."}]},"sound_effects":{"v1":"na","v2":"native","offered":true,"notes":"speak with Seed Audio 1.0 and no voice treats the input as a prompt, which can describe sound effects or ambience. There is no dedicated operation, this applies to that model family only, and no recording exercises it.","sources":["s127"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/speech.rb","method":"render_speech_payload"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak yields SpeechChunk bytes as /audio/speech returns its raw byte stream, then returns the complete Speech.","sources":["s127"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"speak, stream_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"hexgrad/kokoro-82m","spec":"spec/ruby_llm/speech_streaming_spec.rb","example":"RubyLLM::Speech streams speech from openrouter","cassette":"spec/fixtures/vcr_cassettes/speech_streams_speech_from_openrouter.yml","notes":"Binary chunks join to the complete Speech."}]},"task_budgets":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"max_tool_calls and stop_server_tools_when bound the server-tool loop and go through with_provider_options. RubyLLM has no shared API for them.","sources":["s269"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openrouter.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. No route returns intents, topics or sentiment. Decisions questions count under Classification, and Extractions under OCR / document parsing.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. The OpenAPI paths have no vector store route.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Batch API returns text and embeddings only. Video output is not batched.","sources":["s867","s129"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"video_editing":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Video input_references and previous_job_id (\"edit or extend\") go through provider_options on supporting models. animate has no editing argument, and exact transformations are model-specific.","sources":["s281"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/videos.rb","method":"render_video_payload"}]},"video_extension":{"v1":"na","v2":"passthrough","offered":true,"notes":"previous_job_id on /videos names a completed job to edit or extend, on models that support continuation; only the reference describes it. It goes through provider_options, and animate has no extend argument.","sources":["s281"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/videos.rb","method":"render_video_payload"}]},"voice_cloning":{"v1":"missing","v2":"passthrough","offered":true,"notes":"input_references with audio clips clone a voice per request on supporting endpoints. They go through speak provider_options, with no shared voice-reference argument and no voice resource to manage.","sources":["s127"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter/speech.rb","method":"render_speech_payload"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. ","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited inference, server-tool, media, Files and Batch inventory documents no operation for this. Fetching exists only as a server tool during generation.","sources":["s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents HTTP and SSE. The Codex Desktop guide says to leave Responses WebSocket support off for OpenRouter.","sources":["s876","s274","s856"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/openrouter.rb","method":"protocol"}]}},"gaps":[{"notes":"Chat batches, Embedding batches, Responses batches: RubyLLM posts to /api/beta/batches while the docs document /api/v1/batches, and no batch has completed through a live recording. Image and file URLs on chat and Responses batches raise. callback_url and provider.only are not exposed. OpenRouter has no cancel operation.","sources":["s867","s129"]},{"notes":"Apply patch tool: OpenRouter validates the diff and the application applies it. RubyLLM declares the tool but has no execution or apply_patch_call_output lifecycle.","sources":["s380"]},{"notes":"Tool search: Deferred tools on Responses send a bare tool_search tool, while the docs and schema list openrouter:tool_search. RubyLLM has no alias for it and no recording covers the route.","sources":["s272"]},{"notes":"OCR / document parsing: The alpha Extractions route returns typed fields from images and PDFs, and RubyLLM has no integration for it.","sources":["s864"]},{"notes":"File listing and deletion: RubyLLM implements upload, metadata lookup and download, not GET /files or DELETE /files/{id}.","sources":["s279"]},{"notes":"Guardrail configuration: RubyLLM does not create or assign workspace guardrail policies.","sources":["s863"]},{"notes":"Advisor tool, Task budgets: These need a raw tool hash and request fields. RubyLLM has no alias or budget API for the server-tool loop.","sources":["s868","s269"]},{"notes":"Fast inference: service_tier goes through provider options, with no shared speed API.","sources":["s278"]},{"notes":"JSON object mode: response_format json_object goes through provider options.","sources":["s854"]},{"notes":"Omni video generation, Video editing, Video extension: input_references and previous_job_id go through provider options.","sources":["s281"]},{"notes":"Prefix completion: Chat#complete sends nothing after a trailing assistant message, so the messages array goes through provider options.","sources":["s851"]},{"notes":"Provider-managed fallback: The models list and routing preferences go through provider options.","sources":["s282"]},{"notes":"Response caching: The cache header goes through with_headers.","sources":["s130"]},{"notes":"Voice cloning: Reference clips go through speech provider options.","sources":["s127"]},{"notes":"Image generation: The API accepts n up to 10. RubyLLM rejects count: and reads the first image only.","sources":["s126"]},{"notes":"Context editing, Manual context compaction: the Messages context_management field and the Responses compaction_trigger item work upstream, and RubyLLM has no route for them.","sources":["s869","s274"]}]},{"id":"perplexity","name":"Perplexity","notes":"Covers the Agent API, which is the default chat route and runs statelessly, plus embeddings, generated files, the Orchestrator API (formerly Router, in private preview) and legacy Sonar, whose support ended on September 27, 2026 while synchronous requests keep working. Search, Decisions, Batch and Files are documented with no RubyLLM integration.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"Runs on the Agent API with effort levels or provider/model ids. Sonar stays reachable with protocol: :chat_completions, and Orchestrator with protocol: :router_chat_completions.","sources":["s263","s877"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality perplexity/openai/gpt-5.6-luna can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_perplexity_openai_gpt-5_6-luna_can_have_a_basic_conversation.yml","notes":"POST /v1/agent with store:false returned a completed answer."},{"status":"passed","recorded_on":"2026-10-11","model":"fast","spec":"spec/ruby_llm/providers/perplexity_spec.rb","example":"RubyLLM::Providers::Perplexity Agent API answers through a preset, reporting the billed cost","cassette":"spec/fixtures/vcr_cassettes/providers_perplexity_agent_api_answers_through_a_preset_reporting_the_billed_cost.yml","notes":"The fast preset answered and reported its cost and one web search."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Streams Agent API server-sent events. Function calls arrive whole, without argument deltas, and only a finished function_call item becomes a tool call. Perplexity reports its own searches as function calls that finish as search_results.","sources":["s878","s263"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"build_item_added_chunk, build_item_done_chunk"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses perplexity/openai/gpt-5.6-luna supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_perplexity_openai_gpt-5_6-luna_supports_streaming_responses.yml","notes":"Server-sent events from /v1/agent were assembled into chunks."},{"status":"passed","recorded_on":"2026-10-11","model":"fast","spec":"spec/ruby_llm/providers/perplexity_spec.rb","example":"RubyLLM::Providers::Perplexity Agent API streams through a preset without running its web searches as tools","cassette":"spec/fixtures/vcr_cassettes/providers_perplexity_agent_api_streams_through_a_preset_without_running_its_web_searches_as_tools.yml","notes":"The search Perplexity streams as a function call does not run as a local tool."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema returns parsed JSON Schema output on the Agent API. RubyLLM sends text.format, while Perplexity documents response_format; a recording with text.format returned schema-shaped JSON. protocol: :router_chat_completions sends response_format with strict schemas by default.","sources":["s878","s263","s412"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload, schema_format"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/providers/perplexity_spec.rb","example":"RubyLLM::Providers::Perplexity Agent API returns JSON Schema output through with_schema","cassette":"spec/fixtures/vcr_cassettes/providers_perplexity_agent_api_returns_json_schema_output_through_with_schema.yml","notes":"The request carried text.format with a JSON Schema and the parsed Ruby hash matched."}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"The Agent API renders images as input_image, as a data URI or HTTPS URL. Support varies by model. Sonar also accepts images through protocol: :chat_completions. No image inference is recorded.","sources":["s879","s880"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity/media.rb","method":"render_content"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocol_supported_attachment_spec.rb","example":"RubyLLM::Protocol#supported_attachment? RubyLLM::Protocols::Perplexity::Agent matches rendering support for png attachments","notes":"The Agent protocol renders PNG attachments."}]},"pdf_input":{"v1":"native","v2":"native","offered":true,"notes":"The default Agent API takes PDFs and other documents as input_file, and the legacy Sonar route through protocol: :chat_completions takes them as file_url. Perplexity documents PDF, DOC, DOCX, TXT and RTF.","sources":["s881","s880","s882"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_document"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity/media.rb","method":"render_document"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.4-mini","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models perplexity/openai/gpt-5.4-mini understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_perplexity_openai_gpt-5_4-mini_understands_pdfs.yml","notes":"Inline PDF as input_file on the Agent API."}]},"audio_input":{"v1":"missing","v2":"unknown","offered":null,"notes":"The Orchestrator Chat Completions schema lists an input_audio content part with WAV or MP3 data, and RubyLLM renders it through protocol: :router_chat_completions. No guide describes it and the published Orchestrator catalog names no audio model. The Agent API input schema and Sonar content parts have no audio. The Orchestrator API returned 403 limited preview to this key, so availability could not be tested and remains unverified.","sources":["s412","s258","s314","s881","s263"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"input_audio"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity/media.rb","method":"supported_attachment?"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"serializes documented WAV audio input while keeping Sonar audio unsupported","notes":"Checks WAV encoding only; no audio inference is recorded."},{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /orchestrator/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"Sonar rejects video_url content parts with 400: video_url content blocks are not supported by v2 /chat/completions. The Agent API has no video input. RubyLLM rejects video attachments on every Perplexity route.","sources":["s265","s883","s877","s881"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity/media.rb","method":"supported_attachment?"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"supported_attachment?"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"sonar","notes":"POST /chat/completions with a video_url part returned 400: video_url content blocks are not supported by v2 /chat/completions."}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Agent API output has no reasoning items, so it returns no thinking text; reasoning tokens still land in usage. Sonar Deep Research returns think tags on the legacy route, parsed into Message#thinking through protocol: :chat_completions. No recording covers the Sonar path.","sources":["s263","s884","s882"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity/chat.rb"},{"version":"v1","path":"lib/ruby_llm/providers/perplexity/media.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_think_tags"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_completion_body"}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking renders reasoning.effort on the Agent API, with minimal through max. Models and effort levels decide what applies. reasoning_effort is sent on the Orchestrator and Sonar routes.","sources":["s263","s412","s884"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}]},"citations":{"v1":"passthrough","v2":"native","offered":true,"notes":"Agent API search_results items become Citation objects, streaming included. Sonar root citations and search_results parse on protocol: :chat_completions.","sources":["s263","s878","s877"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"parse_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_root_citations"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"fast","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with perplexity/fast returns search result citations","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_perplexity_fast_returns_search_result_citations.yml","notes":"search_results items became Citation objects."}]},"tools":{"v1":"partial","v2":"native","offered":true,"notes":"Ruby function tools run on the default Agent API. Results replay as function_call_output from local history. The Orchestrator route renders them too, with unit coverage only.","sources":["s375","s263","s267","s412"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"build_item_done_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling perplexity/openai/gpt-5.6-luna can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_perplexity_openai_gpt-5_6-luna_can_use_tools.yml","notes":"A function_call item ran locally and its output was replayed in the next request."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"executes and replays local tool results with reasoning context and normalized cache usage","notes":"Local tool results replay and cache counters normalize."},{"status":"unavailable","checked_on":"2026-11-10","model":"perplexity/kimi-k3","notes":"POST /router/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview. Contact api@perplexity.ai to request access. No successful inference or cache hit is recorded."}]},"tool_choice":{"v1":"partial","v2":"native","offered":true,"notes":"with_tool_options(choice:) renders auto, none, required and a named function on the Agent API and through protocol: :router_chat_completions. No live recording covers it; the Orchestrator request returned HTTP 403.","sources":["s263","s412","s261"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"renders required and named tool choices, parallel controls, schema and cache boundaries through Chat","notes":"Public Chat request rendering for the Orchestrator Chat Completions route."},{"status":"unavailable","checked_on":"2026-11-10","model":"perplexity/kimi-k3","notes":"POST /router/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview. Contact api@perplexity.ai to request access. No successful inference or cache hit is recorded."}]},"parallel_tools":{"v1":"partial","v2":"native","offered":true,"notes":"Several function_call items in one Agent API response run through the Chat tool loop and replay together.","sources":["s375","s263"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_completion_body"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/approvals.rb","method":"parse_pending_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling perplexity/openai/gpt-5.6-luna can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_perplexity_openai_gpt-5_6-luna_can_use_parallel_tool_calls.yml","notes":"The first response carried two function_call items."}]},"parallel_tool_control":{"v1":"partial","v2":"unknown","offered":true,"notes":"with_tool_options(calls:) maps to parallel_tool_calls. Orchestrator documents and honors it. The Agent API request schema does not list it and treats it as an effort-level setting, though RubyLLM sends it there too, so treat it as Orchestrator-only. Orchestrator, formerly Router, is in private preview and requires protocol: :router_chat_completions. RubyLLM posts to the /router/v1 base, which Perplexity keeps working alongside /orchestrator/v1. The configured account gets HTTP 403, so no successful Orchestrator inference is recorded. Unverified until Perplexity grants Orchestrator access: on November 10, 2026 the configured key still got HTTP 403, \"The Orchestrator API is currently in limited preview.\"","sources":["s412","s261","s263","s885"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"renders required and named tool choices, parallel controls, schema and cache boundaries through Chat","notes":"Public Chat request rendering for the Orchestrator Chat Completions route."},{"status":"unavailable","checked_on":"2026-11-10","model":"perplexity/kimi-k3","notes":"POST /router/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview. Contact api@perplexity.ai to request access. No successful inference or cache hit is recorded."}]},"server_web_search":{"v1":"native","v2":"native","offered":true,"notes":"Effort levels search on their own, and with_provider_tools(:web_search) adds the tool to a provider/model request. Searches are counted in server_tool_use. Results are not replayed because Perplexity rejects them as input.","sources":["s262","s263"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/perplexity.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"parse_server_tool_use, render_assistant_items"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"fast","spec":"spec/ruby_llm/chat_citations_spec.rb","example":"RubyLLM::Chat citations with perplexity/fast returns search result citations","cassette":"spec/fixtures/vcr_cassettes/chat_citations_with_perplexity_fast_returns_search_result_citations.yml","notes":"search_results items became Citation objects."}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_fetch) sends the Agent API's fetch_url tool on stateless requests. Results come back as ServerToolCall and usage counts them as web_fetch.","sources":["s262","s263","s267"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"SERVER_TOOL_ALIASES, SERVER_TOOL_USAGE_NAMES"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat perplexity hosted tools fetches a public page and counts the fetch","cassette":"spec/fixtures/vcr_cassettes/chat_perplexity_hosted_tools_fetches_a_public_page_and_counts_the_fetch.yml","notes":"A fetch_url_results item, the answer from the fetched page, and web_fetch_requests of 1."}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:code_execution) sends the Agent API's sandbox tool on stateless requests. The run comes back as a ServerToolCall holding the code and its output, usage counts it as code_execution, and later turns rebuild from the local transcript without resending the sandbox_results item, which Perplexity rejects as input.","sources":["s377","s262","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"SERVER_TOOL_ALIASES, SERVER_TOOL_USAGE_NAMES, render_assistant_items"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat perplexity hosted tools runs code in its sandbox and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_perplexity_hosted_tools_runs_code_in_its_sandbox_and_continues_from_the_local_transcript.yml","notes":"A sandbox_results item, the right product, code_execution_requests of 1, and a second turn built from local history."},{"status":"passed","recorded_on":"2026-11-10","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat perplexity hosted tools streams a sandbox run","cassette":"spec/fixtures/vcr_cassettes/chat_perplexity_hosted_tools_streams_a_sandbox_run.yml","notes":"Streamed text matches the final message, with a sandbox_results item."}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The Agent API tools overview and Tool schema list no file_search type. The inherited RubyLLM alias has nothing to reach. Documents go to the model as input_file, without a search tool.","sources":["s262","s265","s881","s264","s258","s260","s266"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(mcp: { url:, name: }) renders the documented mcp tool with server_url and server_label, and calls run without approval by default. Approvals are documented with previous_response_id and RubyLLM drops mcp_approval_request items on replay, so require_approval is unverified. Connectors take raw definitions. No recording covers MCP.","sources":["s378","s886","s262","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"render_assistant_items"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"Agent API effort levels use stable cache keys automatically, and with_caching(key:) sends prompt_cache_key. Cache reads and writes normalize into Tokens on the Agent API and Orchestrator.","sources":["s885","s882","s268","s314","s412"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"parse_usage"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"prompt_cache_params"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/agent_spec.rb","example":"counts cache writes apart from fresh input","notes":"Separates cache writes, reads and fresh input."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"executes and replays local tool results with reasoning context and normalized cache usage","notes":"Local tool results replay and cache counters normalize."}]},"cache_boundaries":{"v1":"missing","v2":"unknown","offered":true,"notes":"cache_until_here renders prompt_cache_breakpoint with explicit mode on the Orchestrator route. The Agent API documents no breakpoints. Orchestrator Messages cache_control is not reachable. Orchestrator, formerly Router, is in private preview and requires protocol: :router_chat_completions. RubyLLM posts to the /router/v1 base, which Perplexity keeps working alongside /orchestrator/v1. The configured account gets HTTP 403, so no successful Orchestrator inference is recorded. Unverified until Perplexity grants Orchestrator access: on November 10, 2026 the configured key still got HTTP 403, \"The Orchestrator API is currently in limited preview.\"","sources":["s412","s261","s314"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"prompt_cache_breakpoint"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"renders required and named tool choices, parallel controls, schema and cache boundaries through Chat","notes":"Public Chat request rendering for the Orchestrator Chat Completions route."},{"status":"unavailable","checked_on":"2026-11-10","model":"perplexity/kimi-k3","notes":"POST /router/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview. Contact api@perplexity.ai to request access. No successful inference or cache hit is recorded."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"No request control is documented. The xhigh effort level compacts context inside its own run as a fixed setting outside the request schema. with_compaction is dropped with a debug log. Orchestrator Messages accepts context_management and ignores it.","sources":["s885","s265","s266","s413","s264","s258","s260"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"apply_compaction"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"token_counting":{"v1":"unknown","v2":"unknown","offered":null,"notes":"The Orchestrator Messages specification has CountMessageTokensRequest and Response components describing POST /orchestrator/v1/messages/count_tokens, but its paths list only /orchestrator/v1/messages and the Orchestrator guide omits token counting. The Agent API and Chat Completions references have no such endpoint. POST /orchestrator/v1/messages/count_tokens returned 403 limited preview to this key, so availability could not be tested and remains unverified. RubyLLM has no route for it.","sources":["s266","s413","s261","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /orchestrator/v1/messages/count_tokens returned 403: The Orchestrator API is currently in limited preview."}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The image_search tool and Sonar image results retrieve existing images. The Agent, Orchestrator and Sonar references document no image generation or editing operation.","sources":["s262","s880","s265","s258","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The image_search tool and Sonar image results retrieve existing images. The Agent, Orchestrator and Sonar references document no image generation or editing operation.","sources":["s262","s880","s265","s258","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"Sonar can return found videos, which is retrieval. The Agent, Orchestrator and Sonar references document no video generation operation.","sources":["s880","s265","s258","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The references document no moderation operation. Safety identifiers attribute policy violations to an end user and do not classify content.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.embed posts to /v1/embeddings and decodes base64 signed int8 vectors, with custom dimensions. Binary quantization is not decoded.","sources":["s115","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity/embeddings.rb","method":"embedding_url, parse_embedding_response, decode_embedding"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"pplx-embed-v1-0.6b","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding Perplexity int8 embeddings perplexity/pplx-embed-v1-0.6b decodes a single text into int8 vectors","cassette":"spec/fixtures/vcr_cassettes/embedding_perplexity_int8_embeddings_perplexity_pplx-embed-v1-0_6b_decodes_a_single_text_into_int8_vectors.yml","notes":"POST /v1/embeddings returned base64 vectors that decode to signed int8."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The Embeddings API takes text only: standard and contextualized embeddings have no image input.","sources":["s115","s265","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"No rerank endpoint is documented. A cookbook ranks sentences through Decisions API probabilities, which is application code over another operation.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"file_upload":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.upload posts to /v1/files, which stores JSONL batch input only and requires purpose: \"batch\". Other files get the provider's own error. Agent API requests do not read uploaded file ids.","sources":["s265","s887","s881"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/files.rb","method":"files_url"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/files.rb","method":"render_upload_fields, parse_file_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with perplexity/openai/gpt-5.6-luna answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_perplexity_openai_gpt-5_6-luna_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"A real JSONL upload to /v1/files with purpose=batch, used by the batch that follows."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM::UploadedFile.find and RubyLLM.download take a response_id/files/file_id identifier for files an Agent API run produced. A bare file id reads batch input and output through /v1/files/{file_id}/content. No recording covers a generated file; batch output downloads in the recorded batch.","sources":["s881","s265","s377","s887"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/files.rb","method":"find, parse_response_file, split_resource_id, download_file_url"},{"version":"v2","path":"lib/ruby_llm/uploaded_file.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/files_spec.rb","example":"finds and downloads only the bound generated file through the provider connection","notes":"Metadata lookup and byte download keep the response-scoped identifier."}]},"chat_batches":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.batch uploads Agent API requests as JSONL and posts /v1/batches, with one model per batch on a 24 hour window, for every organization except zero data retention ones. Function tools are rejected by the provider at upload. The create sends an Idempotency-Key so a 429 can be retried: the lowest usage tier allows one batch or file request per second.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol, batch_protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/batches.rb","method":"create_batch, batch_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with perplexity/openai/gpt-5.6-luna answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_perplexity_openai_gpt-5_6-luna_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"Two chats submitted, polled to completed, and both answers appended to their conversations."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"Perplexity Batch API accepts only /v1/responses lines, so it has no batch for this operation. Multiple inputs to embed run synchronously.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"agent_responses":{"v1":"missing","v2":"native","offered":true,"notes":"The default chat protocol, run statelessly: store:false, the whole conversation each turn, no previous_response_id. Covers effort levels, provider/model ids, streaming, function tools, JSON Schema output, the web_search tool and search_results citations. Replayed turns drop search_results and fetch_url_results, which Perplexity rejects as input. Cost uses the reported total, which includes tool fees. RubyLLM selects levels through the deprecated preset field, and its list omits ultra. Documents are rejected before sending. Stored responses and previous_response_id chaining stay outside RubyLLM coverage.","sources":["s263","s267","s262","s882","s885","s881"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"PRESETS, render_payload, render_assistant_items, parse_usage"},{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"agent_url, protocol_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"fast","spec":"spec/ruby_llm/providers/perplexity_spec.rb","example":"RubyLLM::Providers::Perplexity Agent API continues a searched conversation","cassette":"spec/fixtures/vcr_cassettes/providers_perplexity_agent_api_continues_a_searched_conversation.yml","notes":"The second turn replays local history, without previous_response_id."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/agent_spec.rb","example":"does not run the web searches Perplexity streams as function calls","notes":"Pins the streamed search handling."}]},"agent_skills":{"v1":"missing","v2":"passthrough","offered":true,"notes":"The Agent API skills field takes built-in, inline and custom skills, and with_provider_options(skills:) forwards it. Selecting a skill enables the sandbox tool. RubyLLM does not manage /v1/skills resources.","sources":["s265","s263","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"async_research":{"v1":"missing","v2":"partial","offered":true,"notes":"Perplexity ended asynchronous Sonar requests when Sonar support ended on September 27, 2026, and points them to Agent API background mode. with_provider_options(background: true) can submit a run, but RubyLLM has no polling, retrieval or cancellation, so the job cannot be completed.","sources":["s877","s888","s884"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"background_responses":{"v1":"missing","v2":"partial","offered":true,"notes":"POST /v1/agent takes background, with GET /v1/agent/{id} and cancel. The request field passes through with_provider_options, but RubyLLM has no polling, retrieval or cancellation lifecycle and sends store:false.","sources":["s888","s265","s267"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"completion_url"}]},"browser_use":{"v1":"na","v2":"unknown","offered":null,"notes":"The Working with Files page and Agent API schema say browser tools are unavailable with documents, including when an effort level enables one, but the tools overview and Tool schema list no browser tool. Live, the Agent API accepts tools: [{ type: \"browser\" }], where a nonsense type is rejected, and openai/gpt-5.4-nano replied in two runs that no browser tool was available. Whether it is a usable request feature remains unverified.","sources":["s881","s262","s265","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"openai/gpt-5.4-nano","notes":"POST /v1/agent with a browser tool was accepted and returned a text reply with no browser call."}]},"classification":{"v1":"na","v2":"missing","offered":true,"notes":"POST /v1/decisions answers yes/no, choice and score questions with probabilities, for classifying and routing. It takes text, JSON or images. Perplexity registers no judgment protocol, so RubyLLM.judge cannot reach it.","sources":["s889","s265","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/decisions.rb","method":"render_judgment_payload"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"Orchestrator Messages accepts context_management, and its contract says it is ignored and context edits are not applied. The Agent API has no such control.","sources":["s266","s413","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/contextualizedembeddings takes nested document chunks and returns nested groups. RubyLLM.embed posts only to /v1/embeddings.","sources":["s890","s115","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity/embeddings.rb","method":"embedding_url"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"fast_inference":{"v1":"na","v2":"passthrough","offered":true,"notes":"service_tier: \"priority\", or its alias \"fast\", selects higher-priority processing on supported Agent API models, and Orchestrator honors it too. It goes through with_provider_options; RubyLLM has no dedicated switch.","sources":["s891","s412","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"file_management":{"v1":"na","v2":"missing","offered":true,"notes":"/v1/files lists, retrieves and deletes batch files, and GET /v1/agent/{id}/files lists a response's files. RubyLLM exposes no listing or deletion for Perplexity.","sources":["s265","s887","s881"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/files.rb","method":"find"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Google Maps grounding. It is not a generic capability claim about this provider.","sources":["s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"previous_response_id continues a server-side conversation, and store:false only hides a response from retrieval. RubyLLM keeps conversation state in the application and replays it.","sources":["s267","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/agent.rb","method":"render_assistant_items"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"Perplexity Batch API accepts only /v1/responses lines, so it has no batch for this operation.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Interactions API. It is not a generic capability claim about this provider.","sources":["s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"json_mode":{"v1":"partial","v2":"passthrough","offered":true,"notes":"Orchestrator documents response_format json_object. with_provider_options forwards it on protocol: :router_chat_completions. The Agent API response_format enum lists only json_schema. Orchestrator, formerly Router, is in private preview and requires protocol: :router_chat_completions. RubyLLM posts to the /router/v1 base, which Perplexity keeps working alongside /orchestrator/v1. The configured account gets HTTP 403, so no successful Orchestrator inference is recorded.","sources":["s412","s258","s261","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"leaves request controls and tool descriptions to Router","notes":"with_provider_options values reach the Orchestrator payload. JSON object output has no recording."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Bedrock Mantle protocols. It is not a generic capability claim about this provider.","sources":["s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Perplexity Batch API accepts only /v1/responses lines, so it has no batch for this operation.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Agent API and Orchestrator models all use one Perplexity request format per endpoint, with no publisher-specific wire format to select. The Responses API row covers the unregistered Orchestrator formats.","sources":["s261","s891","s258","s260","s266"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"passthrough","offered":true,"notes":"Code the model runs in the sandbox can call Web Search, Fetch URL and People Search through a preinstalled Perplexity SDK. This takes the raw { type: \"sandbox\" } tool. Application functions are not callable from sandbox code.","sources":["s377","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}]},"prompt_cache_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_caching(key:, ttl:, mode:) renders prompt_cache_key and prompt_cache_options through protocol: :router_chat_completions. The Agent API accepts prompt_cache_key in its frozen configurations, and RubyLLM sends it there too. Orchestrator Messages cache_control is not reachable. Orchestrator, formerly Router, is in private preview and requires protocol: :router_chat_completions. RubyLLM posts to the /router/v1 base, which Perplexity keeps working alongside /orchestrator/v1. The configured account gets HTTP 403, so no successful Orchestrator inference is recorded.","sources":["s412","s258","s885","s261"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"prompt_cache_params, build_prompt_cache_options"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"prompt_cache_params"},{"version":"v2","path":"lib/ruby_llm/protocols/perplexity/router.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/perplexity/router_spec.rb","example":"renders required and named tool choices, parallel controls, schema and cache boundaries through Chat","notes":"Public Chat request rendering for the Orchestrator Chat Completions route."},{"status":"unavailable","checked_on":"2026-11-10","model":"perplexity/kimi-k3","notes":"POST /router/v1/chat/completions returned 403: The Orchestrator API is currently in limited preview. Contact api@perplexity.ai to request access. No successful inference or cache hit is recorded."}]},"provider_server_fallback":{"v1":"missing","v2":"passthrough","offered":true,"notes":"The Agent API models field takes a fallback chain of up to five models, and with_provider_options(models: [...]) forwards it; it takes precedence over the model RubyLLM sends. Orchestrator failover is automatic and takes no setting. with_fallbacks in RubyLLM is a separate feature.","sources":["s892","s265","s268"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"responses":{"v1":"missing","v2":"missing","offered":true,"notes":"Orchestrator documents a stateless /orchestrator/v1/responses endpoint in private preview, and RubyLLM registers no protocol for it. The Agent API has its own row. Orchestrator Messages is also unreachable.","sources":["s260","s261","s266"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"native","offered":true,"notes":"Every Perplexity batch line targets /v1/responses, which is the Agent API, so RubyLLM.batch is this operation.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol, batch_protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/batches.rb","method":"batch_endpoint"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"openai/gpt-5.6-luna","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with perplexity/openai/gpt-5.6-luna answers staged chats and appends the answers to their conversations","cassette":"spec/fixtures/vcr_cassettes/batch_with_perplexity_openai_gpt-5_6-luna_answers_staged_chats_and_appends_the_answers_to_their_conversations.yml","notes":"The recorded batch runs on /v1/responses."}]},"search_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /search returns ranked web results as JSON, without a generated answer. RubyLLM has no operation for it; search inside chat is the Web search tool row.","sources":["s893","s265","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Agent API tool list has image_search for finding images and no generation tool. The inherited RubyLLM alias has nothing to reach.","sources":["s262","s265","s258","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"native","offered":true,"notes":"defer_loading: true on the mcp alias lets the model search that server's tools, and results come back as tool_search_output items. Connectors are deferred by default. It covers remote MCP tools only: the Tool schema has no tool_search type for deferred Ruby tools, which RubyLLM would add. No recording covers it.","sources":["s378","s886","s265","s263"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES, tool_search?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_tool_search"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The Tool schema has no shell or computer tool. The model runs bash inside the sandbox, which the Code execution tool row covers.","sources":["s262","s377","s265","s258","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"task_budgets":{"v1":"missing","v2":"passthrough","offered":true,"notes":"max_steps bounds the Agent API research loop, up to 500, and with_provider_options(max_steps:) forwards it. A provider/model request defaults to one step. RubyLLM has no dedicated budget setting.","sources":["s265","s894","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"Perplexity Batch API accepts only /v1/responses lines, so it has no batch for this operation.","sources":["s887","s265"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"No standalone fetch endpoint is documented. fetch_url is an Agent API tool, and URL fetching in the Decisions API serves that decision only.","sources":["s265","s262","s889","s264"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The Agent, Orchestrator, Search, Embeddings, Decisions, Batch, Files and legacy Sonar references document no operation for this. Perplexity Computer and arbitrary code running in a sandbox are not counted as separate inference endpoints.","sources":["s264","s265","s258","s260","s266","s882"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/perplexity.rb","method":"protocol"}]}},"gaps":[{"notes":"File listing and deletion: /v1/files lists, retrieves and deletes batch files. RubyLLM has no listing or deletion for them.","sources":["s265","s887"]},{"notes":"Asynchronous research, Background Responses jobs: background: true can be sent, but there is no polling, retrieval or cancellation. Perplexity ended asynchronous Sonar on September 27, 2026.","sources":["s888","s877"]},{"notes":"Classification: The Decisions API returns probabilities for yes/no, choice and score questions. RubyLLM.judge has no Perplexity protocol.","sources":["s889"]},{"notes":"Contextual embeddings: /v1/contextualizedembeddings has no renderer or parser.","sources":["s890"]},{"notes":"Standalone search API: POST /search returns ranked results with no RubyLLM operation.","sources":["s893"]},{"notes":"Responses API: The stateless Orchestrator Responses endpoint, and Orchestrator Messages, have no registered protocol. Both are in private preview.","sources":["s260","s266","s261"]},{"notes":"Programmatic tool calling: code the model runs in the sandbox can call Perplexity hosted tools through a preinstalled SDK, and application functions are not callable from it.","sources":["s262","s377"]},{"notes":"Agent skills, Task budgets, Fast inference, Provider-managed fallback, JSON object mode: skills, max_steps, service_tier, models and json_object go through with_provider_options.","sources":["s265","s891","s892","s412"]},{"notes":"Perplexity Agent API: RubyLLM selects effort levels through the deprecated preset field. Its list omits the ultra level, and profiles and pinned versions have no dedicated setting.","sources":["s882","s885"]},{"notes":"Remote MCP tools: require_approval is documented with previous_response_id, and RubyLLM drops approval items on replay, so the approval flow is unverified.","sources":["s378"]},{"notes":"Cache boundaries, Prompt cache options, Limit parallel calls: Orchestrator is in private preview and the configured account gets HTTP 403, so these rest on unit coverage.","sources":["s261","s412"]},{"notes":"Audio input, Token counting: documented only for the Orchestrator API, which returned 403 limited preview to this key, so availability remains unverified. Browser use: the Agent API accepts a browser tool, and openai/gpt-5.4-nano did not use it.","sources":["s258","s883","s266","s881"]}]},{"id":"deepgram","name":"Deepgram","notes":"This covers speech generation, transcription, Text Intelligence and Deepgram LLMs on /v1/think, which is in Early Access. RubyLLM integrates Aura speech, Nova transcription including streaming output, and chat on /v1/think. Voice Agent conversations are outside this release; their documented offerings remain listed separately. Account, project and billing administration are outside scope.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.chat with the Deepgram provider posts to /v1/think on agent.deepgram.com, in Early Access. The application's history goes out in full on every request. think-fast and phonellm-mini-alpha-2-fp8 are not in the model registry and need assume_model_exists: true, because Deepgram's model listing covers speech and transcription only. Every Think-Fast request is stored by Deepgram for model training and abuse monitoring. The endpoint rejects max_tokens and other fields outside its list, and RubyLLM reports that error instead of dropping them.","sources":["s895","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol_for, think_url"},{"version":"v2","path":"lib/ruby_llm/providers/deepgram/chat.rb","method":"completion_url, render_role"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload, parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint answers from the history the application holds","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_answers_from_the_history_the_application_holds.yml","notes":"A second turn recalls a name that only the local history carries."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"A block on ask streams chat.completion.chunk events, always with the stream_options the endpoint requires. Usage arrives on the final chunk.","sources":["s895","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint streams chunks and reports usage","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_streams_chunks_and_reports_usage.yml","notes":"Chunks join into the final content and usage is reported."}]},"structured_output":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint rejects response_format as an unknown field. Voice Agent documents no schema-constrained output.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"vision":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint requires content to be a plain string and rejects content parts. Voice Agent takes audio and text messages only.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint requires content to be a plain string and rejects content parts. Voice Agent documents no document input.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"audio_input":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Deepgram Voice Agent takes live audio in a conversation API. This is outside RubyLLM 2.0 coverage. The stateless /v1/think endpoint accepts plain string content only.","sources":["s231","s232","s233","s372","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint requires content to be a plain string. Voice Agent documents no video input.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"thinking":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"The Voice Agent AgentThinking message carries the text of the agent thought in its content field. This is outside RubyLLM 2.0 coverage. The /v1/think format has no reasoning field: it reports reasoning_tokens, and phoneLLM mini puts reasoning text in the reply.","sources":["s897","s233","s896","s898"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort. These models are not in the registry, so an effort is required: use effort: :none to turn thinking off, because with_thinking(false) and a bare with_thinking raise ArgumentError. The docs say to leave Think-Fast unset or at none, but it accepted low and high live and reported reasoning tokens. On phoneLLM mini any effort other than none puts the reasoning in the reply, followed by a lone closing think tag with no opening tag, so RubyLLM cannot separate it from the answer.","sources":["s895","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint thinks when asked to and not when told not to","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_thinks_when_asked_to_and_not_when_told_not_to.yml","notes":"effort low reports reasoning tokens; effort none reports zero."},{"status":"passed","recorded_on":"2026-11-10","model":"phonellm-mini-alpha-2-fp8","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint accepts a thinking effort on phoneLLM mini","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_accepts_a_thinking_effort_on_phonellm_mini.yml","notes":"The request with reasoning_effort low succeeds."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"with_tools sends function tools, and each turn resends the tool calls and results from the local history. tool_choice is rejected by the endpoint.","sources":["s896","s895"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint runs a tool and sends the result back","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_runs_a_tool_and_sends_the_result_back.yml","notes":"A tool runs and the answer uses its result."},{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint runs a tool from a streamed response","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_runs_a_tool_from_a_streamed_response.yml","notes":"The tool loop also runs over a stream."}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint rejects tool_choice as an unknown field. Voice Agent function settings have no equivalent control.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"A recording shows several tool calls in one response from Think-Fast. The docs type tool_calls as an array and do not describe this behaviour.","sources":["s896","s900"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint takes several tool calls in one response","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_takes_several_tool_calls_in_one_response.yml","notes":"One response carries a weather call for each of two cities."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint accepts eight top-level fields and none limits parallel calls. Voice Agent function settings have no equivalent control.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_web_search":{"v1":"na","v2":"na","offered":false,"notes":"On /v1/think, function is the only supported tool type. Voice Agent functions run on the client or on an endpoint the application supplies, so they are not hosted tools.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"On /v1/think, function is the only supported tool type. Voice Agent functions run on the client or on an endpoint the application supplies, so they are not hosted tools.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"On /v1/think, function is the only supported tool type. Voice Agent functions run on the client or on an endpoint the application supplies, so they are not hosted tools.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"On /v1/think, function is the only supported tool type. Voice Agent functions run on the client or on an endpoint the application supplies, so they are not hosted tools.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"On /v1/think, function is the only supported tool type, and the WebSocket specification has no MCP setting. The documented Deepgram MCP server is a CLI feature for coding tools, not a model tool.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"prompt_caching":{"v1":"na","v2":"native","offered":true,"notes":"Deepgram caches repeated prompts by itself and has no request option, so with_caching sends nothing. Repeating a long prompt returned cached_tokens on both models, though not on every request. RubyLLM counts them as cache reads. The docs list the usage field and do not describe caching.","sources":["s896"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/providers/deepgram/chat.rb","method":"openai_prompt_caching?"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"think-fast","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram chat on the Think endpoint reports the cached part of a repeated prompt","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_chat_on_the_think_endpoint_reports_the_cached_part_of_a_repeated_prompt.yml","notes":"Repeats of a long prompt report cached tokens as cache_read."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint rejects unknown fields and requires string content, so it has no cache markers.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The inventories have no cache resource endpoints.","sources":["s235","s233","s896","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"Voice Agent agent.think.context_length limits the characters retained in context, when a custom think endpoint is used. This is outside RubyLLM 2.0 coverage. The stateless /v1/think endpoint has no such control.","sources":["s233","s372","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"The inventories have no token counting endpoint. /v1/think reports token usage only after generation.","sources":["s235","s896","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak reaches Aura voices on POST /v1/speak and returns a complete Speech. Flux voices are served only on /v2/speak, including its stateless batch REST route, and RubyLLM always posts to /v1/speak.","sources":["s101","s901","s902","s903","s235"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/speech.rb","method":"speak, speech_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"aura-2-thalia-en","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram audio speaks text with aura-2","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_audio_speaks_text_with_aura-2.yml","notes":"POST /v1/speak returns audio/mpeg and a typed Speech."}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts local bytes or a remote URL to /v1/listen and normalizes text, words and utterance segments. The prompt becomes keyterm, and provider_options join the query string.","sources":["s100","s235"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/transcription.rb","method":"transcribe, transcription_params"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"nova-3-general","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram audio transcribes a local file with nova-3","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_audio_transcribes_a_local_file_with_nova-3.yml","notes":"Local audio through /v1/listen with words and utterances."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe streams Nova transcription over the authenticated /v1/listen WebSocket, yielding revisable partials, committed timed segments and a completed Transcription with word timestamps. Input is an existing audio file. Caller-fed microphone input and turn-based Flux on /v2/listen, which Deepgram builds for voice conversations, are separate paths RubyLLM does not reach.","sources":["s103","s904","s233"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/transcription.rb","method":"transcribe"},{"version":"v2","path":"lib/ruby_llm/protocols/deepgram.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/streaming_transcription.rb"}],"verification":[{"status":"passed","recorded_on":"2026-09-17","model":"nova-3-general","spec":"spec/ruby_llm/transcription_websocket_spec.rb","example":"RubyLLM::Transcription streams transcription through deepgram WebSockets","cassette":"spec/fixtures/websocket_cassettes/transcription_deepgram.json","notes":"Recorded file exchange with five Results events and one Metadata event, partial and committed text, a typed final transcription and word timing."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/deepgram/streaming_transcription_spec.rb","notes":"Covers partial and committed segments, word speakers, and sending all audio before CloseStream."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"Passing speaker_names sends diarize_model=latest on both /v1/listen routes, and words carry numeric speaker indexes. Deepgram does not name speakers, so RubyLLM rejects non-empty speaker_names and speaker_references.","sources":["s102","s100","s103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/transcription.rb","method":"transcription_hint_params, DIARIZE_MODEL, rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"nova-3-general","spec":"spec/ruby_llm/providers/deepgram_spec.rb","example":"RubyLLM::Providers::Deepgram audio labels words with speakers when speaker names are given","cassette":"spec/fixtures/vcr_cassettes/providers_deepgram_audio_labels_words_with_speakers_when_speaker_names_are_given.yml","notes":"The request sends diarize_model=latest and the recorded words carry speaker values."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The speech APIs have profanity filtering and redaction, but the inventory has no content-moderation classification endpoint. Filtering transcription output is not counted as RubyLLM.moderate coverage.","sources":["s235","s236","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"Listen accepts audio bytes or URLs and Speak returns generated audio, but the REST inventory has no managed file-resource API. Audio input and output are counted in their own rows.","sources":["s235"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"Listen accepts audio bytes or URLs and Speak returns generated audio, but the REST inventory has no managed file-resource API. Audio input and output are counted in their own rows.","sources":["s235"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider API: the Perplexity Agent API. It is not a generic capability claim about this provider.","sources":["s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"classification":{"v1":"missing","v2":"missing","offered":true,"notes":"Text Intelligence /v1/read exposes intent, topic and sentiment analysis, with custom intents and topics. RubyLLM does not implement the Read API; this is separate from transcription keyword formatting.","sources":["s235","s905","s105"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"Voice Agent context controls set the prompt, load history, inject messages or turn off history. None clears individual context items, and /v1/think takes the full history from the caller. The retained-character limit is recorded under Context compaction.","sources":["s237","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe uses the dedicated /v1/listen prerecorded endpoint.","sources":["s235","s100"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"The inventories have no fast or priority tier. The priority parameter on /v2/speak has one value, low, for asynchronous callback requests.","sources":["s235","s903","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"Listen accepts audio bytes or URLs and Speak returns generated audio, but the REST inventory has no managed file-resource API. Audio input and output are counted in their own rows.","sources":["s235"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Google Maps grounding. It is not a generic capability claim about this provider.","sources":["s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"Voice Agent keeps context during the live connection. Reconnection requires the caller to replay agent.context history, /v1/think is stateless, and the REST inventory stores agent configuration, not conversations that can be retrieved and continued.","sources":["s237","s235","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Interactions API. It is not a generic capability claim about this provider.","sources":["s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint rejects response_format as an unknown field. Voice Agent documents no JSON output mode.","sources":["s896","s233","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Bedrock Mantle protocols. It is not a generic capability claim about this provider.","sources":["s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Deepgram uses batch to mean a prerecorded request. Listen, Speak and Read take a per-request callback URL, but the inventory has no batch job resource.","sources":["s235","s100","s901"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"provider_server_fallback":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"Voice Agent accepts an array of think providers as an ordered fallback chain, retried per request. This is outside RubyLLM 2.0 coverage. RubyLLM model fallbacks are a separate framework feature.","sources":["s906","s233"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Deepgram documents realtime conversations through Voice Agent, with Flux on /v2/listen and /v2/speak for conversational pipelines. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s372","s104","s904","s907","s233"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/think endpoint follows the Chat Completions structure. The inventories document no Responses API.","sources":["s896","s235","s233"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventories document no Responses API or batch job resource.","sources":["s896","s235","s233"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventories have no standalone web search endpoint. The search parameter on Listen finds terms in audio.","sources":["s235","s100","s234"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak yields typed SpeechChunk bytes from the chunked /v1/speak response for Aura voices and returns the complete Speech. Binary HTTP transport validates status and content type, preserves binary encoding and refuses retries after audio delivery. The /v1/speak and /v2/speak WebSockets take text as it is produced, for conversational pipelines, and RubyLLM does not use them.","sources":["s101","s340","s907"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/speech.rb","method":"speak, stream_speech_response"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"stream_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"aura-2-thalia-en","spec":"spec/ruby_llm/speech_streaming_spec.rb","example":"RubyLLM::Speech streams speech from deepgram","cassette":"spec/fixtures/vcr_cassettes/speech_streams_speech_from_deepgram.yml","notes":"Recorded /v1/speak exchange replayed through a block on speak; the joined chunks match the complete Speech."}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"text_intelligence":{"v1":"missing","v2":"missing","offered":true,"notes":"The separate /v1/read endpoint for text sentiment, topics, intents and summaries is absent from RubyLLM. Listen accepts the same analysis options through provider_options, but RubyLLM does not parse those results.","sources":["s905","s235","s105"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/deepgram.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/deepgram/transcription.rb","method":"parse_transcription_response"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The Deepgram REST and WebSocket inventories and the Deepgram LLMs reference cover Listen, Speak, Read, Think and Voice Agent, including agent settings and function messages. They do not document this operation or control. This is an API-scope finding, not a claim about what third-party LLMs attached to Voice Agent can do.","sources":["s234","s235","s233","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"Listen, Speak and Voice Agent use WebSockets, but the inventories document no Responses API to carry over them.","sources":["s233","s235","s896"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/deepgram.rb","method":"protocol"}]}},"gaps":[{"notes":"Chat, Chat streaming, Function tools, Multiple tool calls, Thinking controls, Prompt caching: Deepgram LLMs are in Early Access and their ids are not in the model registry, so they need assume_model_exists: true. On phoneLLM mini an effort other than none puts the reasoning in the reply, which RubyLLM cannot separate from the answer.","sources":["s895","s896"]},{"notes":"Speech generation: Flux voices are served only on /v2/speak. RubyLLM.speak always posts to /v1/speak, so it reaches Aura voices only.","sources":["s901","s902"]},{"notes":"Classification, Text analysis API: Text Intelligence /v1/read exposes sentiment, topic, intent and summary analysis. RubyLLM does not implement the Read API or parse the same options on Listen.","sources":["s905","s235"]}]},{"id":"elevenlabs","name":"ElevenLabs","notes":"Covers ElevenAPI speech and transcription, Image & Video (Flows) generation and assets, and the ElevenAgents API. RubyLLM integrates speech, transcription and Flows media; ElevenAgents conversations and hosted-agent administration stay outside this release, and account and billing administration are out of scope.","features":{"chat":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides chat through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s238","s239","s240","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"streaming":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides chat streaming through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s240","s238","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"structured_output":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"vision":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides image input through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s242","s381","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"pdf_input":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides pdf input through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s242","s381","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"audio_input":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides audio input through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s238","s239","s240","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"thinking":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides thinking results through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s243","s244","s240"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"thinking_controls":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"ElevenLabs exposes reasoning effort, thinking budget and summary settings on hosted agents and workflows. Those settings are not per-request inference controls; hosted-agent configuration and conversations are outside RubyLLM 2.0 coverage.","sources":["s243","s244","s419","s421","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"citations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents offers a source_attribution setting on hosted agents, and conversation transcripts report the knowledge base chunks and documents used. This is available only through hosted conversations, which are outside RubyLLM 2.0 coverage.","sources":["s244","s386","s245"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"tools":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides function tools through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s238","s239","s240","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"tool_choice":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"The only documented tool_choice field is on the Realtime session configuration, which follows the OpenAI Realtime wire protocol. Hosted agent configuration documents no tool choice setting. Realtime conversations are outside RubyLLM 2.0 coverage.","sources":["s909","s910","s244"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"parallel_tools":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides multiple tool calls through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s244","s245","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"parallel_tool_control":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents offers enable_parallel_tool_calls in the hosted agent prompt configuration, and the Realtime session configuration accepts parallel_tool_calls. Both apply to hosted or live conversations, which are outside RubyLLM 2.0 coverage.","sources":["s244","s909","s911"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_web_search":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides web search tool through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s247","s241","s420","s374","s419"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_code_execution":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides code execution tool through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s248","s240","s382"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"server_file_search":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides file search tool through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s246","s241","s385","s240"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"server_mcp":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides remote mcp tools through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s249","s241","s240","s374"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"prompt_caching":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides prompt caching through a live or hosted conversation API. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s243","s245","s382"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"compaction":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents limits active context on hosted chat agents with max_files_in_memory, which replaces older uploaded files with a brief summary. The OpenAPI document also defines compaction settings that no agent schema references. Both apply only to hosted conversations, which are outside RubyLLM 2.0 coverage.","sources":["s242","s419"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"paint submits to /v1/flows/image, polls the documented job states with a bounded timeout and returns an Image with its MIME type and signed content URL. size maps to an aspect ratio. Each request produces one image and count is rejected. The model listing omits image models, so calls use assume_model_exists. The committed recording returned HTTP 402 because the API requires a Pro plan, so only protocol specs exercise a successful response.","sources":["s250","s352","s353"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/images.rb","method":"render_image_payload, parse_image_response, wait_for_image, rejected_options"},{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/media.rb","method":"parse_generation_status"},{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}],"verification":[{"status":"unavailable","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","cassette":"spec/fixtures/vcr_cassettes/protocols_elevenlabs_flows_generates_an_image_through_elevenlabs_image_and_video.yml","checked_on":"2026-10-11","notes":"The recorded call returned HTTP 402: the Image & Video API requires a Pro plan or above. No successful output was reached.","model":"gemini-3.1-flash-lite-image"},{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","checked_on":"2026-11-06","notes":"Payload, polling, failures, timeout and single submission are covered with stubbed responses."}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"paint with: serializes one or more source images as inline_base64 or same-provider asset references, and mask: serializes a mask for the model to accept or reject. The API documents mask on GPT Image models. Generation-ID references go through provider_options. Only protocol specs exercise this path.","sources":["s250","s354","s353"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/images.rb","method":"render_image_payload, validate_paint_inputs!"},{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/media.rb","method":"render_media_reference"}],"verification":[{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","checked_on":"2026-11-06","notes":"Inline, asset and mask references are covered with stubbed responses."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate and animate_later return a VideoJob and a typed Video with its MIME type and raw response. Up to two with: images map to first and last frames. Mixed image, audio or video attachments map to reference arrays, and Creatify Aurora takes one image and one audio attachment without a prompt. Image-only reference lists and generation references go through provider_options. The committed recording returned HTTP 402 because the API requires a Pro plan, so only protocol specs exercise a successful response.","sources":["s251","s355","s354","s353"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/videos.rb","method":"render_video_payload, render_video_inputs, render_character_animation, parse_video_job, parse_video_job_status, download_video"},{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/media.rb","method":"render_media_reference"}],"verification":[{"status":"unavailable","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","cassette":"spec/fixtures/vcr_cassettes/protocols_elevenlabs_flows_generates_a_video_through_an_elevenlabs_video_job.yml","checked_on":"2026-10-11","notes":"The recorded call returned HTTP 402: the Image & Video API requires a Pro plan or above. No successful output was reached.","model":"veo-3.1-fast-generate-001"},{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","checked_on":"2026-11-06","notes":"Frames, reference video, Aurora input and job lifecycle are covered with stubbed responses."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak posts to /v1/text-to-speech/{voice_id} with a selected voice and output_format. RubyLLM maps simple names such as mp3 and wav to ElevenLabs formats and passes exact formats through. WAV at 44.1 kHz requires a Pro plan, and the recording of that request returned HTTP 403.","sources":["s912","s341"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/speech.rb","method":"speak, speech_url, render_speech_payload, parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"eleven_v3","spec":"spec/ruby_llm/providers/elevenlabs_spec.rb","example":"RubyLLM::Providers::ElevenLabs audio speaks text with eleven_v3","cassette":"spec/fixtures/vcr_cassettes/providers_elevenlabs_audio_speaks_text_with_eleven_v3.yml","notes":"Returned MP3 audio with the default voice."}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"transcribe posts a file to /v1/speech-to-text and returns text, language, duration and word timestamps. The webhook and multichannel response shapes can be requested through provider_options, but RubyLLM reads only the single-transcript response.","sources":["s107","s913"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/transcription.rb","method":"render_transcription_payload, parse_transcription_response, render_transcription_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"scribe_v2","spec":"spec/ruby_llm/providers/elevenlabs_spec.rb","example":"RubyLLM::Providers::ElevenLabs audio transcribes audio with scribe_v2","cassette":"spec/fixtures/vcr_cassettes/providers_elevenlabs_audio_transcribes_audio_with_scribe_v2.yml","notes":"Multipart /v1/speech-to-text request returned the transcript and language."},{"status":"passed","recorded_on":"2026-10-11","model":"scribe_v2","spec":"spec/ruby_llm/transcription_timestamps_spec.rb","example":"RubyLLM::Transcription requests word timestamps through elevenlabs with the shared keyword","cassette":"spec/fixtures/vcr_cassettes/transcription_requests_word_timestamps_through_elevenlabs_with_the_shared_keyword.yml","notes":"Sends timestamps_granularity=word and returns timed words."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe streams Scribe v2 Realtime over WebSocket, yielding revisable partials, committed text and a final typed Transcription with language, duration and word timestamps. Input is an existing mono PCM or mu-law WAV file, committed in 20-second segments. Caller-fed microphone input and voice activity commits are not implemented, and ElevenAgents sessions are separate.","sources":["s108","s359"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/streaming_transcription.rb","method":"stream_transcription, collect_transcription, process_transcription_event, validate_streaming_transcription"}],"verification":[{"status":"passed","recorded_on":"2026-09-17","model":"scribe_v2_realtime","spec":"spec/ruby_llm/transcription_websocket_spec.rb","example":"RubyLLM::Transcription streams transcription through elevenlabs WebSockets","cassette":"spec/fixtures/websocket_cassettes/transcription_elevenlabs.json","notes":"Recorded WebSocket exchange from a file with partial and committed text, a typed final transcription and word timing."},{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/streaming_transcription_spec.rb","checked_on":"2026-11-06","notes":"Event normalization, commit boundaries, and completion and error behavior are covered."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"File transcription: speaker_names sets diarize and num_speakers, and returned words carry speaker_id. RubyLLM does not map speaker_id values back to the given names. RubyLLM rejects diarization on streaming transcription.","sources":["s107"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/streaming_transcription.rb","method":"validate_streaming_transcription"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"scribe_v2","spec":"spec/ruby_llm/providers/elevenlabs_spec.rb","example":"RubyLLM::Providers::ElevenLabs audio labels words with speakers when speaker names are given","cassette":"spec/fixtures/vcr_cassettes/providers_elevenlabs_audio_labels_words_with_speakers_when_speaker_names_are_given.yml","notes":"Request carries diarize=true and num_speakers=1; words include speaker_id."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"upload and UploadedFile.find map to /v1/assets, with multipart asset and name fields and typed UploadedFile metadata. These are workspace media assets used by Image & Video, not ElevenAgents knowledge-base documents. purpose, expires_in and uri are rejected. The committed recording returned HTTP 402 because the API requires a Pro plan, so only protocol specs exercise a successful response.","sources":["s356","s357","s354","s353"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/assets.rb","method":"files_url, upload_file_field, render_upload_fields, parse_file_response"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"upload, find, rejected_options"}],"verification":[{"status":"unavailable","spec":"spec/ruby_llm/protocols/elevenlabs/assets_spec.rb","cassette":"spec/fixtures/vcr_cassettes/protocols_elevenlabs_assets_uploads_retrieves_and_downloads_an_elevenlabs_image_asset.yml","checked_on":"2026-10-11","notes":"The recorded call returned HTTP 402: the Image & Video API requires a Pro plan or above. No successful output was reached."},{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/assets_spec.rb","checked_on":"2026-11-06","notes":"Upload, lookup and option rejection are covered with stubbed responses."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"download fetches fresh asset metadata, rejects an asset still processing without a URL, and downloads the signed content without sending API credentials. The committed recording returned HTTP 402 because the API requires a Pro plan, so only protocol specs exercise a successful response.","sources":["s357","s354","s353"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/assets.rb","method":"download"}],"verification":[{"status":"unavailable","spec":"spec/ruby_llm/protocols/elevenlabs/assets_spec.rb","cassette":"spec/fixtures/vcr_cassettes/protocols_elevenlabs_assets_uploads_retrieves_and_downloads_an_elevenlabs_image_asset.yml","checked_on":"2026-10-11","notes":"The recorded call returned HTTP 402: the Image & Video API requires a Pro plan or above. No successful output was reached."},{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/assets_spec.rb","checked_on":"2026-11-06","notes":"Download, redirect and processing states are covered with stubbed responses."}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Agent / Responses API. It is not a generic capability claim about this provider.","sources":["s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"agent_skills":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenAgents Procedures are reusable task-specific instructions loaded when relevant, with create/compile/get/delete endpoints. RubyLLM does not integrate these hosted skills/procedures.","sources":["s252","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"classification":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides classification as hosted-agent configuration, knowledge-base administration or post-conversation analysis. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s253","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe uses the dedicated /v1/speech-to-text endpoint.","sources":["s107","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/transcription.rb","method":"transcription_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"scribe_v2","spec":"spec/ruby_llm/providers/elevenlabs_spec.rb","example":"RubyLLM::Providers::ElevenLabs audio transcribes audio with scribe_v2","cassette":"spec/fixtures/vcr_cassettes/providers_elevenlabs_audio_transcribes_audio_with_scribe_v2.yml","notes":"Multipart /v1/speech-to-text request returned the transcript and language."}]},"dubbing":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenLabs documents dubbing through /v1/dubbing. RubyLLM has no dubbing operation.","sources":["s914","s915"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs.rb"}]},"fast_inference":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents offers a fast service_tier on supported LLMs through the hosted agent prompt configuration. It is not a per-request inference control, and hosted-agent configuration is outside RubyLLM 2.0 coverage. Speech and media requests document no equivalent tier.","sources":["s908","s244","s419"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenLabs documents listing and deleting workspace assets. RubyLLM implements upload, metadata lookup and download, with no public list or delete operation. Hosted-agent knowledge-base and conversation-file administration are outside the release scope.","sources":["s916","s917","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/assets.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"list_uris"}]},"file_search_store_management":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides file search store management as hosted-agent configuration, knowledge-base administration or post-conversation analysis. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s246","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"fine_tuning":{"v1":"missing","v2":"missing","offered":true,"notes":"The public music fine-tunes API creates custom music fine-tunes. RubyLLM has no training/job lifecycle integration. This is separate from the already documented voice-cloning APIs.","sources":["s254"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Google Maps grounding. It is not a generic capability claim about this provider.","sources":["s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides guardrail configuration as hosted-agent configuration, knowledge-base administration or post-conversation analysis. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s255","s244"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs documents provider-managed conversation storage or hosted conversation resources. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s245","s382"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Interactions API. It is not a generic capability claim about this provider.","sources":["s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row names another provider-specific API: Bedrock Mantle protocols. It is not a generic capability claim about this provider.","sources":["s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"music_generation":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenLabs documents music generation through /v1/music. RubyLLM has no music operation.","sources":["s110","s918"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs.rb"}]},"nonchat_batches":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenAgents exposes a batch RAG-index computation API. This specialized knowledge-base operation is not the same as batched chat or embedding-vector generation, and RubyLLM does not integrate it.","sources":["s256","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"omni_video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate and animate_later map mixed image, audio and video attachments to the images, audios and videos reference arrays, documented on Seedance models, and leave model support to ElevenLabs. Creatify Aurora takes one image and one audio attachment. ByteDance models are disabled by default for API requests and need approval. An image-only list maps to frames, so image references alone go through provider_options. Only protocol specs exercise this path; no successful Seedance run is recorded.","sources":["s251","s355","s354","s353","s257"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/videos.rb","method":"render_video_references, render_character_animation"}],"verification":[{"status":"unit","spec":"spec/ruby_llm/protocols/elevenlabs/flows_spec.rb","checked_on":"2026-11-06","notes":"Video reference mapping, Aurora input and typed VideoJob results are covered with stubbed responses."}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"provider_server_fallback":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenAgents supports default/custom/disabled model fallback sequences. RubyLLM does not configure these server-side agent fallbacks.","sources":["s243","s919"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"realtime":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs documents realtime conversations through agent WebSockets and Eleven Realtime, which also accepts OpenAI Realtime clients. These are outside RubyLLM 2.0 coverage; conversation state stays in the application. Scribe Realtime transcription is covered by Streaming transcription.","sources":["s374","s911","s910","s108","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"sound_effects":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenLabs documents text to sound effects through /v1/sound-generation. speak targets text to speech only, so RubyLLM has no sound effects operation.","sources":["s920","s921"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"A block on speak yields typed SpeechChunk bytes from the HTTP stream endpoint and returns the complete Speech. Binary transport validates status and content type, preserves binary encoding and refuses retries after audio delivery. The stream endpoint does not list WAV formats. The WebSocket text input endpoint is not integrated.","sources":["s341","s922","s109"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/speech.rb","method":"speak, speech_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"eleven_flash_v2_5","spec":"spec/ruby_llm/speech_streaming_spec.rb","example":"RubyLLM::Speech streams speech from elevenlabs","cassette":"spec/fixtures/vcr_cassettes/speech_streams_speech_from_elevenlabs.yml","notes":"Recorded /stream response; the spec checks that joined chunks match the complete Speech."}]},"task_budgets":{"v1":"missing","v2":"unknown","offered":null,"notes":"Hosted agent configuration has an interaction_budget override with values such as 5_minutes and 1_hour, but no page explains what it limits. That one enum does not establish a budget across an agent loop. Agent configuration and conversation references were read; max_tokens applies to a single generation and is not counted.","sources":["s244","s419","s245"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"text_intelligence":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides text analysis api as hosted-agent configuration, knowledge-base administration or post-conversation analysis. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s253","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"vector_store_management":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"ElevenLabs Agents provides vector store management as hosted-agent configuration, knowledge-base administration or post-conversation analysis. This is outside RubyLLM 2.0 coverage; conversation state stays in the application.","sources":["s246","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"video_editing":{"v1":"missing","v2":"na","offered":false,"notes":"The Create Video Generation reference lists Seedance, Creatify Aurora and Veo request variants with no editing model or edit endpoint, and the quickstart says the API exposes a subset of the app models. Seedance 2.5 Video Edit, Kling Edit and similar models appear on the Image & Video overview without an API request. Seedance reference videos remain reference-conditioned generation, covered by Omni video generation.","sources":["s251","s353","s257","s419"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs/flows/videos.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The Create Video Generation reference documents no extend model or source-video field, and the quickstart says the API exposes a subset of the app models. Seedance 2.5 Video Extend and Gemini Omni Flash Extend appear on the Image & Video overview without an API request.","sources":["s251","s353","s257","s419"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"voice_cloning":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenAPI exposes instant and professional voice cloning and associated training and voice-management endpoints. RubyLLM accepts an existing voice ID for speech but does not create voices.","sources":["s923","s924","s241"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"voice_conversion":{"v1":"missing","v2":"missing","offered":true,"notes":"ElevenLabs documents speech-to-speech voice conversion. RubyLLM has no voice conversion operation.","sources":["s111","s925"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/elevenlabs.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"Reviewed the full ElevenAPI, ElevenAgents and Flows API inventory, including agent configuration, conversation results, media generation and knowledge-base endpoints. No dedicated public operation or request control for this feature is documented in that inventory. Related capabilities available only by writing an arbitrary external function are not counted as a provider operation.","sources":["s241","s419","s908"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/elevenlabs.rb","method":"protocol"}]}},"gaps":[{"notes":"Agent skills: ElevenAgents Procedures are reusable task-specific instructions loaded when relevant, with list, create, get, remove and draft endpoints. RubyLLM does not integrate these hosted procedures.","sources":["s252","s241"]},{"notes":"Dubbing, Sound effects, Music generation: These ElevenLabs audio operations have no RubyLLM operation.","sources":["s914","s920","s110"]},{"notes":"File listing and deletion: ElevenLabs documents listing and deleting workspace assets. RubyLLM implements upload, metadata lookup and download only. Hosted-agent knowledge-base and conversation-file administration are outside the release scope.","sources":["s916","s917"]},{"notes":"Fine-tuning: The public music fine-tunes API creates custom music fine-tunes. RubyLLM has no training or job lifecycle integration. This is separate from the voice-cloning APIs.","sources":["s254"]},{"notes":"Other batch operations: ElevenAgents exposes a batch RAG-index computation API. This knowledge-base operation is not batched chat or embedding-vector generation, and RubyLLM does not integrate it.","sources":["s256","s241"]},{"notes":"Provider-managed fallback: ElevenAgents supports default, custom and disabled model fallback sequences. RubyLLM does not configure these server-side agent fallbacks.","sources":["s243","s919"]},{"notes":"Voice cloning: ElevenAPI exposes instant and professional voice cloning and associated training and voice-management endpoints. RubyLLM accepts an existing voice ID for speech but does not create voices.","sources":["s923","s924"]},{"notes":"Voice conversion: Speech-to-speech voice conversion has no public RubyLLM operation.","sources":["s111"]},{"notes":"Image generation, Video generation, File upload, File download: The committed recordings returned HTTP 402 because the Image & Video API requires a Pro plan, so these paths have protocol specs but no successful recording.","sources":["s353"]},{"notes":"Video generation, Omni video generation: A with: list of only images maps to first and last frames and accepts at most two, so image references alone and generation references go through provider_options. ByteDance models need approval for API use.","sources":["s251","s354","s257"]},{"notes":"Transcription: Webhook and multichannel responses can be requested through provider_options, but RubyLLM reads only the single-transcript response.","sources":["s107","s913"]},{"notes":"Streaming transcription: Input is an existing mono PCM or mu-law WAV file with manual commits. Caller-fed microphone input and voice activity commits are not implemented.","sources":["s108","s359"]},{"notes":"Streaming speech generation: RubyLLM uses the HTTP stream endpoint, which lists no WAV formats. The WebSocket text input endpoint is not integrated.","sources":["s341","s922"]}]},{"id":"ollama","name":"Ollama","notes":"Scope: a self-hosted Ollama server through its OpenAI-compatible routes and /v1/systemone, plus the hosted web search and fetch API, against the docs and released v0.40.15 source. Server version and model decide which features work; Ollama Cloud is a separate provider.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.chat posts to /v1/chat/completions on the configured Ollama server. The model and server version decide which features work.","sources":["s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality ollama/qwen3 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_ollama_qwen3_can_have_a_basic_conversation.yml","notes":"Recorded against a local Ollama server."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"Chat streams over server-sent events with include_usage, so streamed turns report token usage.","sources":["s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"stream_options"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses ollama/qwen3 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_ollama_qwen3_supports_streaming_responses.yml","notes":"Streamed chat through /v1/chat/completions."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema renders response_format with a JSON Schema, which Ollama applies on /v1/chat/completions. Ollama's structured outputs page says its cloud does not support the feature; this cell covers the local server.","sources":["s132","s131","s926"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"response_format,schema_strict"}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Image attachments render as base64 data URIs in image_url parts. Ollama does not accept image URLs, so RubyLLM downloads a remote image and sends its bytes. The model must support vision.","sources":["s131","s927"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"render_image"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gemma4","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models ollama/gemma4 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_ollama_gemma4_can_understand_local_images.yml","notes":"Base64 image sent to a local vision model."}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"The audited Ollama native and compatibility APIs document no PDF chat input, and the Anthropic compatibility page lists PDF document blocks as unsupported. RubyLLM raises UnsupportedAttachmentError for PDF attachments.","sources":["s283","s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"supported_attachment?"}]},"audio_input":{"v1":"missing","v2":"native","offered":true,"notes":"Audio attachments render as input_audio parts with base64 data and a format. Released v0.40.15 parses that part on /v1/chat/completions. The compatibility page does not list it, and only audio-capable models accept it. No Ollama recording covers this path.","sources":["s926","s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama/media.rb","method":"format_content"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"render_content"},{"version":"v2","path":"spec/ruby_llm/providers/ollama/media_spec.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/ollama/media_spec.rb","example":"RubyLLM::Providers::Ollama::Media.render_content encodes audio attachments in the released Ollama input_audio format","notes":"Covers the rendered request only."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video input operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Ollama returns thinking in the reasoning field of /v1/chat/completions responses and deltas, and RubyLLM exposes it as thinking. Thinking-capable models only.","sources":["s929","s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking ollama/qwen3 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_ollama_qwen3_returns_thinking_when_available.yml","notes":"The response carries a reasoning field beside the answer."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking(effort:) renders reasoning_effort, which Ollama maps to model-defined names and compatibility aliases. Ollama has no token budgets, so budget: raises ArgumentError. Which values apply depends on the model.","sources":["s131","s929"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/ollama/chat_spec.rb","example":"RubyLLM::Providers::Ollama::Chat.render_payload raises for a thinking budget","notes":"Effort rendering and the budget error, without a live server."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"The audited Ollama native and compatibility APIs document no citations, and the Anthropic compatibility page lists citations blocks as unsupported.","sources":["s283","s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools render in the Chat Completions format, on tool-capable models.","sources":["s131","s930"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling ollama/qwen3 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_ollama_qwen3_can_use_tools.yml","notes":"Tool call and follow-up turn."}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"The compatibility page marks tool_choice unsupported on /v1/chat/completions and on the Anthropic route, and the v0.40.15 request struct has no such field. RubyLLM serializes a choice through shared Chat Completions code, but Ollama does not act on it.","sources":["s131","s928","s926"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"tool_choice"}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM parses several tool calls from one response. Whether a model returns several depends on the model.","sources":["s930","s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling ollama/qwen3 can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_ollama_qwen3_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"The recorded response holds three calls."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"Ollama's OpenAI-compatible chat request has neither tool_choice nor parallel_tool_calls, in the docs or in the v0.40.15 request struct. No request control limits the number of calls.","sources":["s131","s926"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parallel_tool_calls"}]},"server_web_search":{"v1":"na","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) with protocol: :responses sends Ollama's built-in web_search tool on /v1/responses, and the server runs the search. It needs a signed-in local server on an account with usage credits. The default Chat Completions protocol has no alias, and the cloud API does not offer the tool on /v1/responses. Ollama returns the search call and an answer with no citation annotations.","sources":["s931","s131","s932","s933"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"Responses, protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"qwen3","spec":"spec/ruby_llm/providers/ollama/responses_spec.rb","example":"RubyLLM::Providers::Ollama::Responses searches the web on the Responses protocol and continues the conversation","cassette":"spec/fixtures/vcr_cassettes/providers_ollama_responses_searches_the_web_on_the_responses_protocol_and_continues_the_conversation.yml","notes":"Local Ollama v0.40.7, signed in with usage credits: the response carried a completed web_search_call, and a follow-up turn replayed it."}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated web fetch tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated code execution tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file search tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated remote mcp tools operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Ollama reuses prompts automatically and reports cached tokens as prompt_tokens_details.cached_tokens, which RubyLLM parses into cache_read. Reuse depends on the runner.","sources":["s934","s926"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool attachments ollama/qwen3 returns text and attachments from tools","cassette":"spec/fixtures/vcr_cassettes/chat_tool_attachments_ollama_qwen3_returns_text_and_attachments_from_tools.yml","notes":"The second turn reports nonzero cached_tokens."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"The audited APIs document no cache markers, and the Anthropic compatibility page lists cache_control blocks as unsupported.","sources":["s283","s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated managed cache resources operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"compaction":{"v1":"unknown","v2":"native","offered":true,"notes":"Chat#compact calls POST /v1/responses/compact with protocol: :responses and continues from the returned compaction item. with_compaction has no Ollama wire option: a local server ignores context_management, and the Responses checklist marks truncation unsupported.","sources":["s935","s936","s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"Responses, protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb","method":"compaction_url"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"compact"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"qwen3","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with ollama","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_ollama.yml","notes":"Local Ollama v0.40.7: Chat#compact calls responses/compact, and the next request continues from the compaction item."}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"The audited APIs document no token counting endpoint, and the Anthropic compatibility page lists /v1/messages/count_tokens as unsupported. Responses report usage after generation.","sources":["s283","s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Ollama removed its experimental image generation API on 2026-07-28. Released v0.40.15 registers no image route, and the documentation index lists none.","sources":["s383","s935","s283"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image editing operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video generation operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"speech":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.speak posts to /v1/audio/speech. Without a voice the model speaks with its own default voice. Ollama returns WAV unless asked for pcm, and RubyLLM labels it so. Other formats, speeds and instructions return a 400. Speech requires v0.40.9 or later and a speech model such as Phonon.","sources":["s937","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"render_speech_payload,speech_url"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/speech.rb","method":"render_speech_payload,parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"phonon","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality ollama/phonon can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_ollama_phonon_can_speak.yml","notes":"A call with no voice or format returns WAV from Phonon on Ollama v0.40.15."}]},"transcription":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts to /v1/audio/transcriptions with model, file, language, prompt, temperature and response_format. Ollama accepts WAV uploads, up to 512 MB and 2 hours, on models with audio input. It returns json, text or verbose_json with segments, and ignores timestamp_granularities. No Ollama recording covers this path. RubyLLM has no operation for the separate /v1/audio/translations route.","sources":["s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"render_transcription_payload,parse_transcription_response"}]},"streaming_transcription":{"v1":"na","v2":"native","offered":true,"notes":"A block on transcribe sends stream=true and reads the transcript.text.delta and transcript.text.done events Ollama documents. No Ollama recording covers this path.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"stream_transcription,STREAM_TYPES"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"Ollama documents no speaker labels. verbose_json returns one segment per transcribed piece of audio, without a speaker.","sources":["s131","s283"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated moderation operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated ocr / document parsing operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.embed posts to /v1/embeddings with string or array input and dimensions. No Ollama recording covers this path.","sources":["s131","s938"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb","method":"render_embedding_payload"}]},"multimodal_embeddings":{"v1":"na","v2":"native","offered":true,"notes":"embed(with:) renders text, one image and one audio clip as Ollama's input object on /v1/embeddings, for multimodal embedding models. Ollama's handler refuses video. The embed reference does not describe this shape; the v0.40.4 release notes and the handler source do.","sources":["s939","s935","s940"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/embeddings.rb","method":"embedding_media?, render_embedding_payload, media_embedding_input"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"embeddinggemma-2","spec":"spec/ruby_llm/embedding_spec.rb","example":"RubyLLM::Embedding multimodal embeddings ollama/embeddinggemma-2 embeds an image alongside text","cassette":"spec/fixtures/vcr_cassettes/embedding_multimodal_embeddings_ollama_embeddinggemma-2_embeds_an_image_alongside_text.yml","notes":"Recorded on Ollama 0.40.7; audio returned a vector in the same check."}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated reranking operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file upload operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file download operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated chat batches operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated embedding batches operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated advisor tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated agent / responses api operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated agent skills operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated asynchronous research operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"Ollama documents stateless Responses support only, and the v0.40.15 request type marks background as not supported. No route retrieves a response later.","sources":["s131","s932","s935"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated browser use operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"classification":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.judge posts to /v1/systemone on decision models such as Clef, and choice questions classify text into labels. Images are accepted on the models that take them; other attachments raise. Requires v0.35.0 or later. No Ollama recording covers this path.","sources":["s941","s942"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"resolve_protocol"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/judgments.rb","method":"render_judgment_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/system_one/judgments.rb","method":"TYPES"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/ollama_spec.rb","example":"RubyLLM::Providers::Ollama judgments routes judgments through System One and chats through Chat Completions","notes":"Routing and rendered request only."}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated computer use operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated context editing operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated contextual embeddings operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"dedicated_transcription_api":{"v1":"native","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts to /v1/audio/transcriptions with model, file, language, prompt, temperature and response_format. Ollama accepts WAV uploads, up to 512 MB and 2 hours, on models with audio input. It returns json, text or verbose_json with segments, and ignores timestamp_granularities. No Ollama recording covers this path. RubyLLM has no operation for the separate /v1/audio/translations route.","sources":["s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"render_transcription_payload,parse_transcription_response"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated dubbing operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated fast inference operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file listing and deletion operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file search store management operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"fill_in_middle":{"v1":"missing","v2":"missing","offered":true,"notes":"/api/generate and /v1/completions document a suffix field for fill-in-the-middle models. RubyLLM has no completions operation.","sources":["s286","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"No fine-tuning operation is documented in the audited Ollama native and compatibility APIs. Model-file imports and LoRA loading do not count as training jobs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated google maps grounding operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated guardrail configuration operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"Ollama documents stateless Responses support; stored conversations and previous_response_id are explicitly unsupported.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image batches operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated interactions api operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"The compatibility API lists JSON mode, reached with provider_options: { response_format: { type: 'json_object' } }. with_schema covers structured JSON Schema output.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"response_format"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_request_options_spec.rb","example":"RubyLLM::Chat with params ollama/qwen3 supports response_format param","cassette":"spec/fixtures/vcr_cassettes/chat_with_params_ollama_qwen3_supports_response_format_param.yml","notes":"The raw option reaches the server and the reply parses as JSON."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated bedrock mantle protocols operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"manual_compaction":{"v1":"unknown","v2":"native","offered":true,"notes":"Chat#compact calls POST /v1/responses/compact with protocol: :responses. The returned item replaces the wire context for later requests, while the local transcript keeps every message. The public docs do not describe this route.","sources":["s935","s936","s131"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"Responses, protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/compaction.rb","method":"compaction_url"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"compact"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"qwen3","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with ollama","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_ollama.yml","notes":"Local Ollama v0.40.7: Chat#compact calls responses/compact, and the next request continues from the compaction item."}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated music generation operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated other batch operations operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated omni video generation operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"No partner model protocol selection is documented. The Anthropic-compatible /v1/messages route serves Ollama's own models.","sources":["s283","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"prefix_completion":{"v1":"missing","v2":"passthrough","offered":true,"notes":"A local server places a final assistant message in the model's turn, without a documented parameter, and results vary by model: gemma4 continued it, and qwen3 continued it inside its reasoning, then restarted the visible answer. Add the partial answer with add_message and call generate. RubyLLM has no dedicated prefill interface.","sources":["s286","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"add_message, generate"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"qwen3:latest","notes":"On Ollama 0.40.7, RubyLLM's generate and /api/chat continued a prefilled \"Zaffre, then\" inside the reasoning, followed by a fresh answer with the closing think tag in the content. gemma4:latest replied with a single colour."}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated programmatic tool calling operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated prompt cache options operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated provider-managed fallback operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated realtime sessions operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated response caching operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"responses":{"v1":"missing","v2":"partial","offered":true,"notes":"Ollama offers stateless /v1/responses with streaming, function tools and reasoning. protocol: :responses selects it, and Chat Completions stays the default. Only a plain answer after compaction has been exercised; streaming, function tools and reasoning on this route have not.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"Responses, protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"qwen3","spec":"spec/ruby_llm/chat_compact_spec.rb","example":"RubyLLM::Chat compacts and continues a conversation with ollama","cassette":"spec/fixtures/vcr_cassettes/chat_compacts_and_continues_a_conversation_with_ollama.yml","notes":"One non-streaming answer through /v1/responses on local Ollama v0.40.7."}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated responses batches operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"search_api":{"v1":"missing","v2":"missing","offered":true,"notes":"Hosted POST https://ollama.com/api/web_search requires an Ollama account key. RubyLLM has no wrapper for it.","sources":["s133"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated apply patch tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image generation tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The local server has no hosted tool search: on v0.40.7, Responses requests with 13 deferred functions and a tool_search tool returned direct function calls and no tool_search_call. The docs do not mention it. The v0.40.15 Responses types convert a tool_search declaration to a function tool for client execution. RubyLLM has no integration.","sources":["s131","s932"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"qwen3:latest","notes":"POST /v1/responses with 13 deferred functions and a tool_search tool returned function_call get_weather in all three runs."}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated x search tool operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated shell and computer tools operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated sound effects operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"streaming_speech":{"v1":"na","v2":"native","offered":true,"notes":"A block on speak yields SpeechChunk bytes as the audio arrives, from Ollama's streamed audio/wav or audio/pcm response. No Ollama recording covers this path.","sources":["s937","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"stream_speech"},{"version":"v2","path":"lib/ruby_llm/protocol/binary_streaming.rb","method":"stream_binary"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated task budgets operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated text analysis api operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated vector store management operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video batches operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video editing operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video extension operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice cloning operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice conversion operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"web_fetch_api":{"v1":"missing","v2":"missing","offered":true,"notes":"Hosted POST https://ollama.com/api/web_fetch requires an Ollama account key. RubyLLM has no wrapper for it.","sources":["s133"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated responses over websocket operation is documented in the audited Ollama native and compatibility APIs.","sources":["s283","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb"}]}},"gaps":[{"notes":"Transcription, Streaming transcription, Dedicated transcription API: Ollama accepts WAV uploads only and ignores timestamp_granularities. /v1/audio/translations has no RubyLLM operation.","sources":["s131"]},{"notes":"Fill-in-the-middle completion: /api/generate and /v1/completions accept a suffix. RubyLLM has no completions operation.","sources":["s286","s131"]},{"notes":"JSON object mode: The compatibility API accepts JSON object mode through raw response_format. with_schema exposes structured JSON Schema output.","sources":["s131"]},{"notes":"Responses API: protocol: :responses reaches stateless /v1/responses. Streaming, function tools and reasoning on it are untested.","sources":["s131"]},{"notes":"Standalone search API, Standalone web fetch API: Hosted /api/web_search and /api/web_fetch require an Ollama account key. RubyLLM has no wrapper for them.","sources":["s133"]}]},{"id":"ollama_cloud","name":"Ollama Cloud","notes":"Scope: direct cloud access at https://ollama.com, covering the OpenAI-compatible /v1 API RubyLLM uses, the native /api and Anthropic-compatible /v1/messages routes it does not use, System One, and the hosted search and fetch APIs. Local server features count only where the docs or cloud catalogs place them on the hosted API.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"Cloud models through the Ollama dialect of Chat Completions at https://ollama.com/v1.","sources":["s131","s135"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions, api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-oss:120b","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality ollama_cloud/gpt-oss:120b can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_ollama_cloud_gpt-oss_120b_can_have_a_basic_conversation.yml","notes":"Recorded against ollama.com/v1/chat/completions."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"Server-sent Chat Completions chunks on cloud models, with usage in the final chunk.","sources":["s131","s943"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"cache_read_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"stream_options"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-oss:120b","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses ollama_cloud/gpt-oss:120b supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_ollama_cloud_gpt-oss_120b_supports_streaming_responses.yml","notes":"Recorded against ollama.com/v1/chat/completions."}]},"structured_output":{"v1":"na","v2":"na","offered":false,"notes":"The official guide says Ollama Cloud does not support structured outputs. RubyLLM removes the capability from Cloud models.","sources":["s132"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud/models.rb","method":"model_capabilities"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud model listing reports no structured output, which Ollama Cloud does not support","notes":"Checks the capability list a Cloud model gets."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Cloud models that report vision in the catalog. The endpoint takes base64 images and rejects image URLs; RubyLLM always sends a data URI.","sources":["s131","s927","s291"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"render_image"},{"version":"v2","path":"lib/ruby_llm/attachment.rb","method":"for_llm"}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"Chat Completions documents text and image parts only, and the Anthropic-compatible route lists PDF document blocks as unsupported. RubyLLM raises UnsupportedAttachmentError for a PDF.","sources":["s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"supported_attachment?"}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision. An input_audio part on /v1/chat/completions reached gpt-oss:20b as an omitted image placeholder. RubyLLM sends input_audio parts through the Ollama dialect, which cannot work here.","sources":["s131","s944","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"render_audio"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"gpt-oss:20b","notes":"POST /v1/chat/completions with an input_audio part returned empty content; the reasoning quoted Ollama's image omitted placeholder."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The chat references document text and image parts only. RubyLLM raises UnsupportedAttachmentError for video.","sources":["s131","s944"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/media.rb","method":"supported_attachment?"}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"Cloud models that report thinking return reasoning text, which RubyLLM reads into thinking.","sources":["s131","s929","s291"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-oss:120b","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling ollama_cloud/gpt-oss:120b can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_ollama_cloud_gpt-oss_120b_can_use_tools.yml","notes":"The recorded gpt-oss:120b response carries a reasoning field. The spec does not assert the thinking text."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort; each model defines its values, and unsupported names resolve to the model default. Ollama has no thinking budget, so RubyLLM raises when given one.","sources":["s131","s929"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/ollama/chat_spec.rb","example":"RubyLLM::Providers::Ollama::Chat .render_payload raises for a thinking budget","notes":"Covers the Ollama dialect that OllamaCloud::ChatCompletions inherits."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"No citations operation is documented. The Anthropic-compatible route lists citation blocks as unsupported.","sources":["s135","s283","s946","s131","s947","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"Function tools on cloud models that report tools in the catalog.","sources":["s131","s930","s291"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-oss:120b","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling ollama_cloud/gpt-oss:120b can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_ollama_cloud_gpt-oss_120b_can_use_tools.yml","notes":"Recorded against ollama.com/v1/chat/completions, through the tool result round."}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"The compatibility checklists mark tool_choice unsupported on both /v1/chat/completions and /v1/messages. RubyLLM sends tool_choice when you set it, and the endpoint decides what to do with it.","sources":["s131","s928","s135"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"tool_choice"}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"Ollama documents parallel tool calls, and RubyLLM reads every call in a response. The gpt-oss:120b recordings returned one call per turn, so several calls in one response have no Cloud recording.","sources":["s930","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"The compatibility checklist lists no parallel_tool_calls field, and neither the chat reference nor the Anthropic-compatible route documents a limit on parallel calls.","sources":["s131","s944","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parallel_tool_calls"}]},"server_web_search":{"v1":"na","v2":"na","offered":false,"notes":"Ollama search is a REST API that you call yourself or wrap as a function tool. The cloud API excludes built-in web search through /v1/responses, and the Anthropic-compatible route says hosted web search is not fully supported.","sources":["s133","s131","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"Ollama web fetch is a REST API that you call yourself or wrap as a function tool. No hosted fetch tool is documented.","sources":["s133","s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated code execution tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file search tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated remote MCP tools operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"prompt_caching":{"v1":"unknown","v2":"native","offered":true,"notes":"Automatic, with no RubyLLM switch. Cloud chat responses report prompt_tokens_details.cached_tokens, which RubyLLM reads into cache_read. The cloud usage API counts cached input tokens and pricing lists a cached input rate for most models. The docs do not describe the cache rules.","sources":["s948","s949"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"gpt-oss:120b","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses ollama_cloud/gpt-oss:120b reports token usage with and without streaming","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_ollama_cloud_gpt-oss_120b_reports_token_usage_with_and_without_streaming.yml","notes":"The recorded usage carries cached_tokens. The spec asserts token usage overall, not a cache hit."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"No explicit cache markers are documented. The Anthropic-compatible route lists cache_control blocks as unsupported.","sources":["s135","s283","s946","s131","s947","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated managed cache resources operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"No context compaction control is documented, and the Responses checklist marks truncation unsupported.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"No token counting operation is documented. The Anthropic-compatible route lists /v1/messages/count_tokens as unsupported.","sources":["s135","s283","s946","s131","s947","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image generation operation is documented in the Ollama Cloud API. POST /v1/images/generations returned 404, the same as a nonsense path. RubyLLM raises an error naming the missing operation before it sends anything.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/images/generations on ollama.com returned 404: path \"/v1/images/generations\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for image generation without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image editing operation is documented in the Ollama Cloud API. POST /v1/images/edits returned 404, the same as a nonsense path. RubyLLM raises an error naming the missing operation before it sends anything.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/images/edits on ollama.com returned 404: path \"/v1/images/edits\" not found, the same as a nonsense path."}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video generation operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"Not served on ollama.com: POST /v1/audio/speech returned 404, the same as a nonsense path. None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision. The speech guide shows only local servers. RubyLLM raises an error naming the missing operation before it sends anything. Local Ollama still speaks.","sources":["s131","s937","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/audio/speech on ollama.com returned 404: path \"/v1/audio/speech\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for speech without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"Not served on ollama.com: POST /v1/audio/transcriptions returned 404, the same as a nonsense path. None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision. RubyLLM raises an error naming the missing operation before it sends anything. Local Ollama still transcribes.","sources":["s131","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/audio/transcriptions on ollama.com returned 404: path \"/v1/audio/transcriptions\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for transcription without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"Not served on ollama.com: POST /v1/audio/transcriptions returned 404, the same as a nonsense path. None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision. RubyLLM raises an error naming the missing operation before it sends anything.","sources":["s131","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/audio/transcriptions on ollama.com returned 404: path \"/v1/audio/transcriptions\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for streamed transcription without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The transcription reference documents no speaker fields, and ollama.com answers 404 for transcription itself.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated moderation operation is documented in the Ollama Cloud API. POST /v1/moderations returned 404, the same as a nonsense path. RubyLLM raises an error naming the missing operation before it sends anything.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/moderations on ollama.com returned 404: path \"/v1/moderations\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for moderation without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated OCR / document parsing operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Not served to API keys on ollama.com. POST /v1/embeddings returned 404, the same as a nonsense path. POST /api/embed returned the same 401 for a key that works on /api/chat and /api/show, for a made-up key, for every model tried and for an empty body, and local-only /api/ps answers that key with the same 401. Ollama reports plan limits as 402, so this is a closed route and not a limit on this key, though other credentials, such as the signed device login the Ollama app uses, cannot be tested with a key. None of the 23 models in the cloud catalog reports an embedding capability, and the embeddings guide shows only a local server. RubyLLM raises an error naming the missing operation before it sends anything. Local Ollama still embeds.","sources":["s940","s938","s131","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"nomic-embed-text","notes":"POST /v1/embeddings returned 404: path \"/v1/embeddings\" not found. POST /api/embed returned 401 unauthorized with this key, with a made-up key, and for nomic-embed-text, embeddinggemma, gpt-oss:20b, gemma4:31b and an unknown model."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for embeddings without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated multimodal embeddings operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated reranking operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file upload operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file download operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"No chat batches operation is documented. The Anthropic-compatible route lists the Batches API as unsupported.","sources":["s135","s283","s946","s131","s947","s928"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated embedding batches operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated advisor tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Perplexity Agent API is a Perplexity product. The Ollama Cloud inventory documents no equivalent.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated agent skills operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated asynchronous research operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Responses checklist documents stateless requests with no background field.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated browser use operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"classification":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM.judge reaches System One at POST /v1/systemone on https://ollama.com, with Clef as the cloud model, for choice, yes/no and scoring questions. It needs a paid plan or usage credits.","sources":["s942","s941","s950","s291"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/providers/ollama.rb","method":"SystemOne"},{"version":"v2","path":"lib/ruby_llm/providers/ollama/judgments.rb","method":"judgment_url"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated computer use operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated context editing operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated contextual embeddings operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"Not served on ollama.com: POST /v1/audio/transcriptions returned 404, the same as a nonsense path. None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision.","sources":["s131","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"transcription_url"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/audio/transcriptions on ollama.com returned 404: path \"/v1/audio/transcriptions\" not found, the same as a nonsense path."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated dubbing operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"No fast inference option is documented in the API. Pricing lists Fast mode as coming soon.","sources":["s135","s283","s946","s131","s947","s949"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file listing and deletion operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated file search store management operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"No cloud model reports the insert capability that fill-in-the-middle requires: /api/show gives completion, tools, thinking, vision and decision across the catalog. RubyLLM has no completions operation.","sources":["s131","s286","s291"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"gpt-oss:20b","notes":"/api/show lists no insert capability on any of the 23 cloud models."}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"No fine-tuning job is documented, and creating or deleting models requires a local Ollama server. Model imports do not count as training.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated google maps grounding operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated guardrail configuration operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The cloud API does not support stateful Responses: previous_response_id and conversation are unsupported.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image batches operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"The Ollama Cloud inventory documents no Interactions API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"The structured-output guide excludes Ollama Cloud and covers both JSON mode and schema-constrained output.","sources":["s132"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Bedrock Mantle is a Bedrock route. The Ollama Cloud inventory documents no equivalent.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated manual context compaction operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated music generation operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"No batch operation is documented for other operations.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated omni video generation operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Ollama Cloud documents one set of routes for all its models, with no per-partner wire format.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"passthrough","offered":true,"notes":"gpt-oss:20b continues a final assistant message on /api/chat and /v1/chat/completions; gemma4:31b and nemotron-3-super restart the answer, so support varies by model. Ollama documents no prefix parameter. Add the partial answer with add_message and call generate. RubyLLM has no dedicated prefill interface.","sources":["s131","s286"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"add_message, generate"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"gpt-oss:20b","notes":"/api/chat, /v1/chat/completions and RubyLLM's generate continued a prefilled \"Zaffre, then\" in place. gemma4:31b and nemotron-3-super returned a fresh list."}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated programmatic tool calling operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated prompt cache options operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated provider-managed fallback operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated realtime sessions operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated response caching operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"responses":{"v1":"unknown","v2":"missing","offered":true,"notes":"Cloud documents stateless /v1/responses requests at https://ollama.com/v1, with streaming, function tools and reasoning summaries, and without built-in web search or custom tool replay. OllamaCloud registers only Chat Completions, so RubyLLM cannot select Responses here.","sources":["s131"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated responses batches operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"search_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST https://ollama.com/api/web_search is documented for API keys. RubyLLM has no search operation for it, and it sits outside the chat request, so raw options cannot reach it.","sources":["s133"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated apply patch tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated image generation tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated tool search operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated x search tool operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated shell and computer tools operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated sound effects operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"Not served on ollama.com: POST /v1/audio/speech returned 404, the same as a nonsense path. None of the 23 models in the cloud catalog reports this: capabilities are completion, tools, thinking, vision and decision. RubyLLM raises an error naming the missing operation before it sends anything.","sources":["s131","s937","s291","s945"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"ChatCompletions"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/audio/speech on ollama.com returned 404: path \"/v1/audio/speech\" not found, the same as a nonsense path."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/ollama_cloud_spec.rb","example":"RubyLLM::Providers::OllamaCloud operations ollama.com does not serve raises for streamed speech without a request","notes":"Raises before any request, and no request reaches the stubbed route."}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated task budgets operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated text analysis API operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated vector store management operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video batches operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video editing operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated video extension operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice cloning operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated voice conversion operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"web_fetch_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST https://ollama.com/api/web_fetch is documented for API keys. RubyLLM has no fetch operation for it, and it sits outside the chat request, so raw options cannot reach it.","sources":["s133"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"No dedicated Responses over WebSocket operation is documented in the Ollama Cloud API.","sources":["s135","s283","s946","s131","s947"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/ollama_cloud.rb","method":"protocol"}]}},"gaps":[{"notes":"Responses API: Cloud documents stateless /v1/responses requests; OllamaCloud registers only Chat Completions.","sources":["s131"]},{"notes":"Standalone search API, Standalone web fetch API: The hosted search and fetch REST APIs have no RubyLLM operation.","sources":["s133"]},{"notes":"Speech generation, Streaming speech generation, Transcription, Streaming transcription, Dedicated transcription API: the compatibility reference documents these without limiting them to local servers, but ollama.com answers 404 for the routes, so RubyLLM raises an error before sending a request. Text embeddings: /v1/embeddings answers 404, and /api/embed returns the same 401 to a working key and a made-up key, so RubyLLM raises an error before sending a request.","sources":["s131","s291","s945"]},{"notes":"Fill-in-the-middle completion, Prefix completion: The completions routes document suffix and raw prompts, but no page establishes either behavior on a cloud model. RubyLLM has no completions operation.","sources":["s131","s286"]}]},{"id":"gpustack","name":"GPUStack","notes":"Scope: GPUStack 2.3 standard inference routes and its documented built-in backends, including vLLM and vLLM-Omni. Support depends on the deployed model, backend version and settings, and vLLM or vLLM-Omni routes outside the gateway need Generic Proxy enabled on the deployment. GPU administration and external model-provider integrations are outside the comparison.","features":{"chat":{"v1":"native","v2":"native","offered":true,"notes":"Chat Completions at /v1/chat/completions on deployed text models. Backend and model decide what each deployment accepts.","sources":["s136","s295"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality gpustack/qwen3 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_gpustack_qwen3_can_have_a_basic_conversation.yml","notes":"Successful chat response."}]},"streaming":{"v1":"native","v2":"native","offered":true,"notes":"stream: true on Chat Completions with SSE deltas and final usage.","sources":["s136","s295"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses gpustack/qwen3 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_gpustack_qwen3_supports_streaming_responses.yml","notes":"Streamed deltas from the gateway route."}]},"structured_output":{"v1":"native","v2":"native","offered":true,"notes":"with_schema sends response_format json_schema, which vLLM documents. GPUStack does not document it separately, and no GPUStack recording covers it. Other backends may differ.","sources":["s136","s951"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"response_format"}]},"vision":{"v1":"native","v2":"native","offered":true,"notes":"Images are sent inline as image_url parts on vision models. Local and remote images both become data URLs, because clusters often cannot reach the internet.","sources":["s136","s952","s329"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack/media.rb","method":"render_image"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack_spec.rb","notes":"sends remote images inline, because clusters often cannot reach the internet"}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document PDF input.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"audio_input":{"v1":"missing","v2":"native","offered":true,"notes":"input_audio content parts on audio-capable vLLM models. The deployed model must support audio.","sources":["s136","s329"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack/media.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack/media.rb","method":"render_audio"}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"video_url content parts, with remote and local videos sent as data URLs. Requires a video-capable vLLM or vLLM-Omni deployment; other backends and models may not accept video.","sources":["s292","s329"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack/media.rb"},{"version":"v1","path":"lib/ruby_llm/providers/gpustack/media.rb","method":"format_content"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack/media.rb","method":"render_video"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack/media_spec.rb","notes":"accepts video attachments through the public chat API; sends remote videos inline; encodes local video attachments as data URLs"}]},"thinking":{"v1":"native","v2":"native","offered":true,"notes":"Reads reasoning_content or reasoning from responses and streams, and replays it on later turns. vLLM needs a reasoning parser on the deployment; GPUStack does not document this field separately.","sources":["s953","s136"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking gpustack/qwen3 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_gpustack_qwen3_returns_thinking_when_available.yml","notes":"Response carries reasoning_content."},{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking gpustack/qwen3 streams thinking content when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_gpustack_qwen3_streams_thinking_content_when_available.yml","notes":"Stream carries reasoning_content."}]},"thinking_controls":{"v1":"native","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort, which the GPUStack Playground and some vLLM reasoning models use. Model-specific switches such as chat_template_kwargs enable_thinking go through provider options. Effect depends on the model and backend, and no GPUStack recording sends an effort.","sources":["s953","s295","s954"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"reasoning_effort"}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Citations.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"tools":{"v1":"native","v2":"native","offered":true,"notes":"Function tools on Chat Completions. Deployments need tool calling enabled (for vLLM, --enable-auto-tool-choice and a tool call parser). vLLM and SGLang document it.","sources":["s955","s956"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling gpustack/qwen3 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_gpustack_qwen3_can_use_tools.yml","notes":"Tool call and follow-up turn."}]},"tool_choice":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(choice:) sends tool_choice. vLLM documents auto, required, none and a named function.","sources":["s955","s956"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control gpustack/qwen3 respects choice: :none","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_gpustack_qwen3_respects_choice_none.yml","notes":"Request sends tool_choice none; the recording covers this value only."}]},"parallel_tools":{"v1":"native","v2":"native","offered":true,"notes":"Reads several tool calls from one response. vLLM says the count depends on the model.","sources":["s956","s295"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling gpustack/qwen3 can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_gpustack_qwen3_can_use_parallel_tool_calls.yml","notes":"Response holds several tool calls."}]},"parallel_tool_control":{"v1":"native","v2":"native","offered":true,"notes":"with_tool_options(calls:) sends parallel_tool_calls. vLLM documents false as returning at most one call. No GPUStack recording sends the field.","sources":["s295"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parallel_tool_calls"}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(web_search: { require_approval: \"never\" }) selects the search subtool of the configured web_search_preview server. Requires protocol: :responses and a tool server configured on the vLLM deployment with --tool-server. RubyLLM requires an explicit require_approval: \"never\" because vLLM has no approval events, and it rejects per-request URLs, connectors and authorization because vLLM connects only to its configured servers. The allowed_tools filter limits the tool descriptions shown to the model, so restrict execution on the tool server itself. RubyLLM keeps call records as the backend returns them and never invents a result.","sources":["s293","s957","s958","s136"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb","method":"GPUStack inherits OpenAI Chat Completions; no Responses route"},{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"completion_url fixed to chat/completions; inherited by GPUStack in tag 1.16.0"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"MCP_ALIASES, render_mcp_alias, merge_mcp_filters"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/responses_spec.rb","notes":"maps portable server tools to only their configured vLLM namespace and subtools"}]},"server_web_fetch":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(web_fetch: { require_approval: \"never\" }) selects the open subtool of the configured web_search_preview server. vLLM dispatch outside Harmony calls only search, so open needs a Harmony model such as the gpt-oss family. Requires protocol: :responses and a tool server configured on the vLLM deployment with --tool-server. RubyLLM requires an explicit require_approval: \"never\" because vLLM has no approval events, and it rejects per-request URLs, connectors and authorization because vLLM connects only to its configured servers. The allowed_tools filter limits the tool descriptions shown to the model, so restrict execution on the tool server itself. RubyLLM keeps call records as the backend returns them and never invents a result.","sources":["s293","s957","s958","s136"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb","method":"GPUStack inherits OpenAI Chat Completions; no Responses route"},{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"completion_url fixed to chat/completions; inherited by GPUStack in tag 1.16.0"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"MCP_ALIASES, render_mcp_alias, merge_mcp_filters"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/responses_spec.rb","notes":"combines search and fetch filters so vLLM cannot overwrite one with the other"}]},"server_code_execution":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(code_execution: { require_approval: \"never\" }) selects the configured code_interpreter server only, without container tools. Requires protocol: :responses and a tool server configured on the vLLM deployment with --tool-server. RubyLLM requires an explicit require_approval: \"never\" because vLLM has no approval events, and it rejects per-request URLs, connectors and authorization because vLLM connects only to its configured servers. The allowed_tools filter limits the tool descriptions shown to the model, so restrict execution on the tool server itself. RubyLLM keeps call records as the backend returns them and never invents a result.","sources":["s293","s957","s958","s136"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/gpustack.rb","method":"GPUStack inherits OpenAI Chat Completions; no Responses route"},{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb","method":"completion_url fixed to chat/completions; inherited by GPUStack in tag 1.16.0"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"MCP_ALIASES, render_mcp_alias"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/responses_spec.rb","notes":"maps portable server tools to only their configured vLLM namespace and subtools"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document File search tool.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"server_mcp":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(mcp: ...) reaches the servers configured on vLLM under the labels it dispatches: web_search_preview, code_interpreter and container. Requires protocol: :responses and a tool server configured on the vLLM deployment with --tool-server. RubyLLM requires an explicit require_approval: \"never\" because vLLM has no approval events, and it rejects per-request URLs, connectors and authorization because vLLM connects only to its configured servers. The allowed_tools filter limits the tool descriptions shown to the model, so restrict execution on the tool server itself. RubyLLM keeps call records as the backend returns them and never invents a result.","sources":["s293","s957","s959","s958","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"MCP_LABELS, validate_mcp_tool, render_server_tool_history"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/responses_spec.rb","notes":"sends configured MCP labels and preserves complete calls, usage, and multi-turn replay; rejects approval modes, per-request servers, and ignored read-only filters before HTTP"}]},"prompt_caching":{"v1":"native","v2":"native","offered":true,"notes":"Built-in backends reuse prefixes automatically, and shared cache services extend that for vLLM and SGLang. RubyLLM reads prompt_tokens_details.cached_tokens into cache_read. Deployment settings decide whether caching runs and is reported.","sources":["s960","s961","s294","s962"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/chat.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"qwen3","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses gpustack/qwen3 reports token usage with and without streaming","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_gpustack_qwen3_reports_token_usage_with_and_without_streaming.yml","notes":"Usage includes cached_tokens."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Cache boundaries.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Cache services are shared KV infrastructure that administrators attach to deployments. GPUStack documents no per-request cache resource that clients create.","sources":["s961","s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"compaction":{"v1":"na","v2":"native","offered":true,"notes":"with_compaction on Chat Completions sends vLLM's truncate_prompt_tokens with truncation_side: \"left\", keeping the last at: tokens. With no at:, it keeps as many as the model allows when an output limit is set; with neither, it does not truncate, because vLLM rejects a prompt that leaves no room to answer. This is plain truncation, not summarization: it can cut the system prompt and the earliest turns. instructions: and pause_after: are ignored. vLLM deployments only; on protocol: :responses the option is dropped with a debug log.","sources":["s295"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack/chat.rb","method":"apply_compaction"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"smollm2-135m","spec":"spec/ruby_llm/providers/gpustack/chat_live_spec.rb","example":"RubyLLM::Providers::GPUStack::Chat truncates a long prompt to the compaction threshold and still answers","cassette":"spec/fixtures/vcr_cassettes/providers_gpustack_chat_truncates_a_long_prompt_to_the_compaction_threshold_and_still_answers.yml","notes":"GPUStack 2.2.3, vLLM 0.25.1: a prompt over the 2048-token limit returned 400 without compaction; with at: 40 it used 40 input tokens and answered, and with no at: and an output limit of 8 it used 2040. With no at: and no output limit, vLLM returned 400, so RubyLLM does not truncate in that case."}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#count_tokens and RubyLLM.count_tokens send the rendered messages and tools to vLLM /tokenize and return its count, using a gpustack_api_base ending in /model/proxy/ROUTE_ID/v1 on a route with the model proxy enabled. RubyLLM.tokenize returns exact IDs for plain text on the same route. The gateway base is rejected before any request. vLLM deployments only: pure diffusion servers do not expose /tokenize.","sources":["s963","s964","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/tokenization.rb","method":"count_tokens_url, render_count_tokens_payload, parse_count_tokens_response, tokenization_url, render_tokenization_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"smollm2-135m","spec":"spec/ruby_llm/providers/gpustack/chat_live_spec.rb","example":"RubyLLM::Providers::GPUStack::Chat counts the tokens the next request uses","cassette":"spec/fixtures/vcr_cassettes/providers_gpustack_chat_counts_the_tokens_the_next_request_uses.yml","notes":"GPUStack 2.2.3, vLLM 0.25.1 through the model proxy: the count matched the input tokens the next request reported. Declared tools are covered by a unit spec."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/tokenization_spec.rb","notes":"uses the model proxy tokenizer and preserves IDs and raw backend metadata without charging usage; rejects gateway tokenization before issuing a request"}]},"image_generation":{"v1":"native","v2":"native","offered":true,"notes":"paint uses /v1/images/generations on image deployments. Backend and model support are required.","sources":["s136","s292"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"images_url"}]},"image_editing":{"v1":"native","v2":"native","offered":true,"notes":"paint(with:) uses multipart /v1/images/edits on image deployments. Backend and model support are required.","sources":["s136","s292"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"render_edit_payload"}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate and animate_later use vLLM-Omni /v1/videos with non-idempotent multipart submission, polling, and MP4 download from /content. Requires gpustack_api_base to end in /model/proxy/ROUTE_ID/v1, with Generic Proxy enabled on the deployment; RubyLLM rejects the standard gateway base before any request. GPUStack 2.3 also routes video requests through its gateway, which RubyLLM does not use. One result per call; uploaded file IDs and extension raise before any request. vLLM-Omni also documents video listing and deletion, which RubyLLM does not expose.","sources":["s965","s292","s954","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/videos.rb","method":"video_url, post_video, parse_video_job, download_video"},{"version":"v2","path":"lib/ruby_llm/video_job.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/videos_spec.rb","notes":"submits text-only multipart once, polls all states, and downloads bytes with proxy authentication; rejects unsupported gateway routes, references, multi-output requests, and extension before HTTP"}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak uses /v1/audio/speech on vLLM and VoxBox deployments. Pass voice: and format: explicitly: without them RubyLLM sends the voice alloy, which deployed models may not offer, and labels the result mp3 although vLLM-Omni defaults to wav.","sources":["s136","s296","s966"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"render_speech_payload, parse_speech_response"}]},"transcription":{"v1":"native","v2":"native","offered":true,"notes":"transcribe uses multipart /v1/audio/transcriptions on vLLM and VoxBox deployments. Supported models and optional fields depend on the backend.","sources":["s136","s296","s297"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"render_transcription_payload"}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"transcribe with a block reads vLLM choices[].delta.content, assembles the transcript, and keeps final token usage. Requires the vLLM backend; VoxBox does not stream.","sources":["s296","s297"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack/transcription.rb","method":"stream_transcription, build_transcription_chunk"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack/transcription_spec.rb","notes":"streams vLLM text and retains a separate final usage event; marks the stream done once, with the whole transcript"}]},"diarization":{"v1":"native","v2":"native","offered":true,"notes":"format: \"diarized_json\" returns speaker-labelled segments, and RubyLLM requests it by default for model ids containing diarize. vLLM documents this for MOSS-Transcribe-Diarize only. VoxBox and named speaker references are not documented.","sources":["s297","s296"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb","method":"parse_transcription_response"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"default_response_format, parse_transcription_response"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Moderation.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"ocr":{"v1":"unknown","v2":"na","offered":false,"notes":"GPUStack's inference routes and vLLM-Omni's endpoint list have no OCR or document parsing route. Reading an image through chat is counted under image input.","sources":["s136","s292","s964"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"embeddings":{"v1":"native","v2":"native","offered":true,"notes":"embed uses /v1/embeddings on embedding deployments. Several texts go in one request.","sources":["s136","s967"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/embeddings.rb"},{"version":"v1","path":"lib/ruby_llm/providers/openai/images.rb"},{"version":"v2","path":"lib/ruby_llm/providers/gpustack/embeddings.rb","method":"render_embedding_payload"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack/embeddings_spec.rb","notes":"preserves text batches and provider options"}]},"multimodal_embeddings":{"v1":"missing","v2":"native","offered":true,"notes":"embed(..., with: image) sends vLLM messages and returns one vector. Requires a multimodal embedding model with a chat template on vLLM; several texts cannot be combined with attachments.","sources":["s300","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack/embeddings.rb","method":"embedding_media?, render_embedding_payload"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack/embeddings_spec.rb","notes":"embeds text and an image through the messages request format; accepts image attachments through the public API and returns one vector"}]},"reranking":{"v1":"missing","v2":"native","offered":true,"notes":"rerank uses the Jina-compatible /v1/rerank route on reranker deployments.","sources":["s136","s968"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/rerank.rb","method":"rerank_url, parse_rerank_response"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document File upload.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document File download.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's gateway routes list no batch operation, and a live GPUStack 2.2.3 server answered 404 for POST /v1/chat/completions/batch. vLLM's route of that name, reachable only by enabling Generic Proxy on a vLLM deployment, returns several conversations in one synchronous call: it has no batch id to collect from later, so it is not a batch job. The test server runs llama.cpp, so that vLLM route was not exercised.","sources":["s295","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"POST /v1/chat/completions/batch on the GPUStack 2.2.3 gateway returned 404."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Embedding batches.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Advisor tool.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Perplexity Agent API.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Agent skills.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Asynchronous research.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"background_responses":{"v1":"na","v2":"partial","offered":true,"notes":"vLLM Responses documents retrieval and cancel routes and accepts background when VLLM_ENABLE_RESPONSES_API_STORE=1 is set on the deployment. GPUStack lists only Create Response, so those routes need Generic Proxy. provider_options can send background and store, but RubyLLM has no retrieval, polling or cancel lifecycle, and normal requests send store: false.","sources":["s295","s959","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"store"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Browser use.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"classification":{"v1":"missing","v2":"missing","offered":true,"notes":"vLLM documents /classify for classification models, reachable through an enabled Generic Proxy. RubyLLM has no classification request or typed result.","sources":["s969","s136","s970"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Computer use.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Context editing.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Contextual embeddings.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"dedicated_transcription_api":{"v1":"native","v2":"native","offered":true,"notes":"transcribe uses the dedicated multipart /v1/audio/transcriptions route. Supported models and optional fields depend on the backend.","sources":["s136","s296"],"code":[{"version":"v1","path":"lib/ruby_llm/providers/openai/transcription.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Dubbing.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"No fast tier is documented. vLLM's priority field only orders requests in a deployment configured for priority scheduling.","sources":["s295","s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document File listing and deletion.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document File search store management.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"vLLM documents the suffix parameter of /v1/completions as unsupported, and GPUStack and SGLang document no fill-in-the-middle parameter.","sources":["s295","s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack serves LoRA adapters on deployments but documents no training API.","sources":["s971","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Google Maps grounding.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Guardrail configuration.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"hosted_conversations":{"v1":"na","v2":"out_of_scope","offered":true,"notes":"vLLM Responses can continue through previous_response_id when VLLM_ENABLE_RESPONSES_API_STORE=1 is set on the deployment. This is provider-stored state and stays outside RubyLLM 2.0; conversation history stays in the application.","sources":["s295","s959"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"store"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Image batches.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Interactions API.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"json_mode":{"v1":"passthrough","v2":"passthrough","offered":true,"notes":"vLLM accepts response_format json_object, and provider_options forwards it. RubyLLM exposes schema output separately.","sources":["s295","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Bedrock Mantle protocols.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Manual context compaction.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"music_generation":{"v1":"na","v2":"missing","offered":true,"notes":"vLLM-Omni documents /v1/audio/generate for music on diffusion audio models such as Stable Audio, reachable through an enabled Generic Proxy. RubyLLM has no audio generation operation.","sources":["s972","s964","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Other batch operations.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"omni_video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"animate and animate_later map ordered image, audio and video references to vLLM-Omni multipart fields and return typed VideoJob and Video results. Valid combinations depend on the model; vLLM-Omni says not to combine image_reference with video_reference. Requires gpustack_api_base to end in /model/proxy/ROUTE_ID/v1, with Generic Proxy enabled on the deployment; RubyLLM rejects the standard gateway base before any request.","sources":["s965","s292","s136"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/videos.rb","method":"video_references, validate_animate_inputs!"},{"version":"v2","path":"lib/ruby_llm/video_job.rb"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/videos_spec.rb","notes":"serializes image and audio input and nested backend options as JSON multipart fields; preserves ordered multiple references and sends remote references inline"}]},"partner_model_protocols":{"v1":"na","v2":"missing","offered":true,"notes":"GPUStack documents an Anthropic-compatible /v1/messages route on deployed models, and native forwarding for Anthropic providers through protocol: original. RubyLLM declares only Chat Completions and Responses for this provider, so its Anthropic protocol cannot be selected.","sources":["s136","s954","s973","s974"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb","method":"protocol :chat_completions"}]},"prefix_completion":{"v1":"na","v2":"passthrough","offered":true,"notes":"vLLM documents continue_final_message on Chat Completions to continue a final assistant message, used with add_generation_prompt: false. provider_options forwards both after you add the assistant message. GPUStack does not document it separately.","sources":["s295"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Programmatic tool calling.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"prompt_cache_controls":{"v1":"na","v2":"passthrough","offered":true,"notes":"vLLM documents cache_salt to isolate prefix cache reuse, and provider_options forwards it. with_caching(key:) sends prompt_cache_key, which a route with session affinity can use as its body key. GPUStack documents no TTL or retention option.","sources":["s962","s295","s975"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"prompt_cache_params"}]},"provider_server_fallback":{"v1":"na","v2":"missing","offered":true,"notes":"Administrators set a fallback target on a model route, and requests through RubyLLM use it with no client option. RubyLLM does not configure routes or report which target served a request.","sources":["s975","s976","s954"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"realtime":{"v1":"unknown","v2":"unknown","offered":null,"notes":"vLLM documents a /v1/realtime WebSocket for streaming transcription, and vLLM-Omni documents realtime audio over the same path. GPUStack's gateway and Generic Proxy pages do not say whether WebSockets are forwarded to a deployment. If they are, this stays outside RubyLLM 2.0.","sources":["s136","s970","s297","s964"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Response caching.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"protocol: :responses uses /v1/responses on the gateway or model proxy, and normal chat stays on Chat Completions. Usage and MCP items use the shared Responses parser, with vLLM replay normalization. GPUStack also offers per-endpoint protocol selection. Backend and model support are required.","sources":["s136","s954","s295","s957"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"render_assistant_items, parse_reasoning_summary"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/responses_spec.rb","notes":"requires explicit execution consent and keeps ordinary chat on Chat Completions; reads actual Harmony reasoning content when the provider has no summary"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Responses batches.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Standalone search API.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Apply patch tool.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Image generation tool.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Tool search.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document X search tool.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"shell_computer_tools":{"v1":"na","v2":"unknown","offered":null,"notes":"vLLM's Harmony and tool dispatch code has a container tool that expects a stateful Docker server, and its MCP example lists a container label for gpt-oss. No page documents the tool contract. RubyLLM's mcp alias accepts the container label, but no spec or recording covers it, so availability stays unverified.","sources":["s293","s957"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/responses.rb","method":"MCP_LABELS"}]},"sound_effects":{"v1":"na","v2":"missing","offered":true,"notes":"vLLM-Omni documents /v1/audio/generate for sound effects and ambient audio on diffusion audio models, reachable through an enabled Generic Proxy. RubyLLM has no audio generation operation.","sources":["s972","s964","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"streaming_speech":{"v1":"missing","v2":"native","offered":true,"notes":"speak with a block sends stream: true, defaults to PCM, and returns typed audio chunks. Requires the vLLM backend; VoxBox does not stream. vLLM-Omni's separate WebSocket text streaming is not used.","sources":["s296","s966"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack/speech.rb","method":"stream_speech"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/gpustack/speech_spec.rb","notes":"requests the vLLM binary stream with PCM as its default format"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Task budgets.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Text analysis API.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Vector store management.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Video batches.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"video_editing":{"v1":"missing","v2":"native","offered":true,"notes":"animate(with: source_video) sends vLLM-Omni video_reference through the model proxy, for models that support video-to-video. Remote references are sent inline. Uses the same lifecycle as generation. Requires gpustack_api_base to end in /model/proxy/ROUTE_ID/v1, with Generic Proxy enabled on the deployment; RubyLLM rejects the standard gateway base before any request.","sources":["s965","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/gpustack/videos.rb","method":"video_references"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/protocols/gpustack/videos_spec.rb","notes":"accepts a local video through the public attachment API and sends a JSON reference field"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Video extension.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"voice_cloning":{"v1":"missing","v2":"passthrough","offered":true,"notes":"GPUStack documents Qwen3-TTS reference audio voice cloning, and vLLM-Omni documents ref_audio and ref_text on /v1/audio/speech. speak forwards them through provider_options. vLLM-Omni also has /v1/audio/voices for uploading voices, which RubyLLM does not expose.","sources":["s296","s966"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"render_speech_payload"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Voice conversion.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Standalone web fetch API.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"GPUStack's standard inference routes and documented built-in backends do not document Responses over WebSocket.","sources":["s136","s292"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/gpustack.rb"}]}},"gaps":[{"notes":"Web search tool, Web fetch tool, Code execution tool, Remote MCP tools: these need protocol: :responses and a tool server configured on vLLM. Per-request MCP URLs and approvals do not exist upstream, and allowed_tools does not restrict execution. Fetch needs a Harmony model.","sources":["s293","s957","s958"]},{"notes":"Background Responses jobs: background needs a store enabled on the vLLM deployment, and RubyLLM has no retrieval, polling or cancel lifecycle.","sources":["s295","s959"]},{"notes":"Classification: vLLM /classify has no RubyLLM operation.","sources":["s969"]},{"notes":"Music generation, Sound effects: vLLM-Omni /v1/audio/generate has no RubyLLM operation.","sources":["s972","s964"]},{"notes":"Partner model protocols: GPUStack's Anthropic-compatible /v1/messages route cannot be selected, because RubyLLM declares only Chat Completions and Responses here.","sources":["s136","s973"]},{"notes":"Provider-managed fallback: route fallback targets are set by administrators. Requests use them transparently, but RubyLLM cannot configure them or report which target answered.","sources":["s975","s976"]},{"notes":"Video generation, Omni video generation, Video editing: RubyLLM reaches vLLM-Omni only through the model proxy, not the /v1/videos gateway routing added in GPUStack 2.3, and does not expose listing or deletion.","sources":["s965","s954"]},{"notes":"Speech generation: pass voice: and format: explicitly, because the defaults are alloy and mp3 while vLLM-Omni defaults to wav.","sources":["s966"]},{"notes":"Voice cloning: ref_audio and ref_text go through provider_options, and /v1/audio/voices uploads have no RubyLLM API.","sources":["s296","s966"]},{"notes":"JSON object mode, Prefix completion, Prompt cache options: these use vLLM fields through provider_options (response_format, continue_final_message, cache_salt).","sources":["s295","s962"]},{"notes":"Realtime sessions, Shell and computer tools: GPUStack WebSocket forwarding and vLLM's container tool contract are not documented.","sources":["s970","s957"]}]},{"id":"hetzner","name":"Hetzner","notes":"Scope: Hetzner's OpenAI-compatible Inference API serving Qwen open-weight models, not its separate OpenClaw experiment. Its one documentation page, last changed 2026-07-24, describes an experimental service at /api/v1 with no parameter reference, so streaming, tools and thinking rest on API recordings made at /v1 on 2026-11-04 with Qwen3.8-27B.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"Documented /v1/chat/completions, reached through a Hetzner dialect of Chat Completions. Instructions use the system role.","sources":["s977","s978"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"render_role"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality hetzner/Qwen3.8-27B can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_hetzner_qwen3_8-27b_can_have_a_basic_conversation.yml","notes":"Recorded against https://inference.hetzner.com/v1/chat/completions; asserts the answer, role and token usage."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"The Inference API page does not mention streaming. Recorded responses show the /v1/chat/completions stream works, with usage in the final chunk through stream_options.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"build_chunk"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"stream_options"}],"verification":[{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses hetzner/Qwen3.8-27B supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_hetzner_qwen3_8-27b_supports_streaming_responses.yml","notes":"Recorded streamed response; asserts chunks arrive and the final response carries status 200."}]},"structured_output":{"v1":"na","v2":"native","offered":true,"notes":"with_schema sends a json_schema response_format and Qwen3.8-27B follows it: the page has no parameter reference, but live replies matched a schema that an unprompted answer could not. Invalid response_format types return 400.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload, json_schema"},{"version":"v2","path":"lib/ruby_llm/models.json","method":"structured_output"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"A schema with an enum of two invented values returned only those values, three times through raw requests and once through with_schema, for a prompt that did not ask for JSON."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"The page documents text and image input for its Qwen models. RubyLLM sends every image, including remote ones, as an inline data URL, although the page shows image URLs.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"render_content, render_inline_image"}],"verification":[{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models hetzner/Qwen3.8-27B can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_hetzner_qwen3_8-27b_can_understand_local_images.yml","notes":"Request carries a data:image/png URL; asserts the reply describes the image."},{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models hetzner/Qwen3.8-27B can understand remote images without extension","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_hetzner_qwen3_8-27b_can_understand_remote_images_without_extension.yml","notes":"RubyLLM downloads the image, then sends it inline."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/hetzner_spec.rb","example":"RubyLLM::Providers::Hetzner attachments sends remote images inline, because Hetzner cannot fetch URLs","notes":"Checks the data URL rendered for a remote image."}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"The page lists text and image modalities only. RubyLLM raises UnsupportedAttachmentError for documents before sending.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"supported_attachment?, document_attachments: :none"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/hetzner_spec.rb","example":"RubyLLM::Providers::Hetzner attachments rejects documents and audio, which the models do not accept","notes":"Checks that a PDF raises UnsupportedAttachmentError."}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The page lists text and image modalities only. RubyLLM raises UnsupportedAttachmentError for audio before sending.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"supported_attachment?, audio_attachments: false"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/hetzner_spec.rb","example":"RubyLLM::Providers::Hetzner attachments rejects documents and audio, which the models do not accept","notes":"Checks that a WAV file raises UnsupportedAttachmentError."}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The page lists text and image modalities only. RubyLLM rejects video attachments before sending.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"supported_attachment?"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"supported_attachment?"}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"The page does not document thinking output. Recorded Qwen3.8-27B responses from 2026-11-04 return message.reasoning, delta.reasoning and reasoning_tokens, and RubyLLM parses all three. No spec asserts the thinking text for this provider.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text, thinking_tokens"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"build_chunk"}]},"thinking_controls":{"v1":"na","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort, and the values depend on the model. Qwen3.8-27B takes none, low, medium and xhigh (its default), with no measurable ordering between them on a short prompt, and returns 400 for minimal, high and max, which RubyLLM raises as UnsupportedParameterError on :thinking. Qwen/Qwen3.6-35B-A3B accepts high and max. The page has no parameter reference. A bare call sends nothing, because the registry lists no reasoning options for this model.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort, reasoning_effort"},{"version":"v2","path":"lib/ruby_llm/models.json","method":"reasoning_options"},{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"parse_unsupported_parameter"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"Qwen3.8-27B","spec":"spec/ruby_llm/providers/hetzner_spec.rb","example":"RubyLLM::Providers::Hetzner thinking controls with the live API thinks at the efforts Qwen3.8-27B takes","cassette":"spec/fixtures/vcr_cassettes/providers_hetzner_thinking_controls_with_the_live_api_thinks_at_the_efforts_qwen3_8-27b_takes.yml","notes":"Request sends reasoning_effort xhigh; the reply carries reasoning text and the right answer."},{"status":"passed","recorded_on":"2026-11-09","model":"Qwen3.8-27B","spec":"spec/ruby_llm/providers/hetzner_spec.rb","example":"RubyLLM::Providers::Hetzner thinking controls with the live API names thinking when Qwen3.8-27B rejects an effort","cassette":"spec/fixtures/vcr_cassettes/providers_hetzner_thinking_controls_with_the_live_api_names_thinking_when_qwen3_8-27b_rejects_an_effort.yml","notes":"reasoning_effort high returned 400 naming xhigh, medium and low; RubyLLM raised UnsupportedParameterError with parameter :thinking."},{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"Raw requests: none returned 0 reasoning tokens; low, medium and xhigh returned 44 to 51; minimal, high and max returned 400. On Qwen/Qwen3.6-35B-A3B, high and max returned 200."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document citations. Recorded responses return annotations as null.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_annotations"}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"The page does not document tool calling. Recorded Qwen3.8-27B responses show function tools work on /v1/chat/completions, including tool results, streaming and nested parameters.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for, parse_tool_calls"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/models.json","method":"function_calling"}],"verification":[{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling hetzner/Qwen3.8-27B can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_hetzner_qwen3_8-27b_can_use_tools.yml","notes":"Recorded tool call and follow-up; asserts the reply uses the tool result."}]},"tool_choice":{"v1":"na","v2":"native","offered":true,"notes":"with_tool_options(choice:) sends tool_choice, and the API honors named tools, none and required on Qwen3.8-27B. A nonsense value returned 400, and the error names auto as the other accepted value. The page has no parameter reference. The shared spec skips Qwen3.8-27B because the registry has no tool_choice capability.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"tool_choice"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"build_tool_choice"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"A forced record_note ran on a weather question through raw requests and with_tool_options. tool_choice none returned no call, and required forced a call on a prompt that needed no tool."}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"The page does not document this. Recorded Qwen3.8-27B responses return two calls to different tools and three calls to one tool in a single response.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling hetzner/Qwen3.8-27B can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_hetzner_qwen3_8-27b_can_use_parallel_tool_calls.yml","notes":"First response carries weather and best_language_to_learn calls."},{"status":"passed","recorded_on":"2026-11-04","model":"Qwen3.8-27B","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling hetzner/Qwen3.8-27B can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_hetzner_qwen3_8-27b_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"First response carries three dice_roll calls."}]},"parallel_tool_control":{"v1":"na","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) sends parallel_tool_calls: false, and the API honors it on Qwen3.8-27B. The page has no parameter reference.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parallel_tool_calls"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"Four cities returned 4 tool calls in one response by default and with true, and 1 call per response with false, including through with_tool_options(calls: :one)."}]},"server_web_search":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a hosted web search tool. The Chat Completions protocol registers no server tool aliases.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a hosted web fetch tool. The Chat Completions protocol registers no server tool aliases.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a hosted code execution tool. The Chat Completions protocol registers no server tool aliases.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a hosted file search tool. The Chat Completions protocol registers no server tool aliases.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document remote MCP tools. The Chat Completions protocol registers no server tool aliases.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"prompt_caching":{"v1":"na","v2":"native","offered":true,"notes":"Long prompts are cached automatically, with no request option, and the reported counts are multiples of 1600 tokens. RubyLLM reads cached_tokens and created_cache_tokens as cache reads and writes. Short requests, such as the shared recordings, report 0 for both.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens, cache_write_tokens"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"A 9848 token prompt reported created_cache_tokens 8000 and then cached_tokens 8000 on the next request; a longer one reported 38400 of each. RubyLLM's tokens reported cache_read and cache_write."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Not honored: on a cold cache, a 38909 token prompt with prompt_cache_breakpoint after the first third created and then read the same 36800 cached tokens as one without it, as a 59000 token prompt did earlier, and the API accepts unknown content part fields too. The page does not mention cache markers. RubyLLM sends no boundary on Hetzner, so cache_until_here has no effect there; automatic caching still applies.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"inject_cache_breakpoint, openai_prompt_caching?"},{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"render_content"},{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"openai_prompt_caching?"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"Cold prompts with and without a breakpoint a third of the way in each reported created_cache_tokens 36800, then cached_tokens 36800. An earlier pair at 59000 tokens reported 57600 for both."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document managed cache resources.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document context compaction. RubyLLM has no compaction parameter for this protocol.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"apply_compaction"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a token counting endpoint. Responses report usage after generation, which is a different feature.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"count_tokens, tokenize"}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document image generation. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document image editing. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document video generation.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document speech generation. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document transcription. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document streaming transcription.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document speaker identification.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document moderation. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an OCR or document parsing endpoint.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document text embeddings. RubyLLM inherits this operation from Chat Completions, so a call reaches an endpoint the page does not list.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document multimodal embeddings.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a reranking endpoint.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document file upload. The provider declares no files protocol.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"files?"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document file download. The provider declares no files protocol.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"files?"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document chat batches. The provider declares no batch operations.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"batches?"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document embedding batches. The provider declares no batch operations.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"batches?"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an advisor tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document the Perplexity Agent API, which belongs to another provider.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document agent skills.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document asynchronous research.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document background Responses jobs.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a browser use tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a classification endpoint.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a computer use tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document context editing.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document contextual embeddings.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a transcription API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document dubbing.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a fast inference tier.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document file listing or deletion. The provider declares no files protocol.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"files?"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document file search stores.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"Not served: POST /v1/completions and /api/v1/completions return 404 route not found, although the page lists /v1/completions. RubyLLM has no completions operation.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"completion_url"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"POST /v1/completions with and without a suffix returned 404 route not found."}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document fine-tuning.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document Google Maps grounding.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document guardrail configuration.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document hosted conversations.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document image batches.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an Interactions API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"json_mode":{"v1":"na","v2":"passthrough","offered":true,"notes":"response_format with type json_object is honored on Qwen3.8-27B: a prompt that did not ask for JSON returned valid JSON. The page has no parameter reference. with_provider_options forwards it; RubyLLM has no named option.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"Three requests for \"Name a fruit.\" with json_object returned valid JSON objects, and with_provider_options(response_format:) did the same."}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document Bedrock Mantle protocols, which belong to another provider.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a compaction endpoint.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document music generation.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document batch operations.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document omni video generation.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document partner model protocols.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"prefix_completion":{"v1":"na","v2":"passthrough","offered":true,"notes":"Qwen3.8-27B continues a final assistant message with continue_final_message: true and add_generation_prompt: false, sent through with_provider_options after add_message and generate. The continuation arrives in the reasoning field, with empty content. Without those fields the model answers afresh. The page does not document it, and POST /v1/completions is not served.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"completion_url"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"add_message, generate"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"Four requests with continue_final_message continued \"1. Teal\\n2.\" in place, for example \" Plum\\n3. Ochre\" and \" Marigold\\n3. Indigo\". RubyLLM returned the continuation as thinking with content nil."}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document programmatic tool calling.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"prompt_cache_controls":{"v1":"na","v2":"unknown","offered":null,"notes":"The page does not mention cache options. RubyLLM sends no prompt_cache_key or prompt_cache_options to Hetzner, whose dialect turns them off. Requests carrying them are accepted, but the API also accepts unknown fields and caching is automatic, so a key's effect could not be observed either way.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"apply_prompt_cache_params, openai_prompt_caching?"},{"version":"v2","path":"lib/ruby_llm/providers/hetzner/chat.rb","method":"openai_prompt_caching?"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"Qwen3.8-27B","notes":"With and without prompt_cache_key, repeated requests reported the same cached_tokens."}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document provider-managed fallback.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document realtime sessions.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document response caching.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a Responses API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document Responses batches.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a search API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an apply patch tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an image generation tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a hosted tool search.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"tool_search?"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document an X search tool.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document shell or computer tools.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document sound effects.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document streaming speech generation.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document task budgets.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a text analysis API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document vector stores.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document video batches.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document video editing.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document video extension.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document voice cloning.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document voice conversion.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document a web fetch API.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"Hetzner's Inference API page lists /v1/models, /v1/completions and /v1/chat/completions as its only endpoints and does not document Responses over WebSocket.","sources":["s977"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/hetzner.rb","method":"protocol :chat_completions"}]}},"gaps":[{"notes":"Prompt cache options: Hetzner's page has no parameter reference, the API accepts unknown fields, and caching is automatic, so the effect of a key could not be observed. RubyLLM sends neither option nor boundary. Cache boundaries are accepted with no visible effect; POST /v1/completions is not served.","sources":["s977"]}]},{"id":"meta","name":"Meta","notes":"Covers the Meta Model API at api.meta.ai/v1 with Muse Wonder, Muse Spark, Muse Image, and Muse Voice Transcribe. RubyLLM declares its Responses format, the OpenAI Files protocol for /v1/files, and Muse Voice Transcribe on the buffered and event-stream transcription paths. Muse Code, the open-weight models, and SAM segmentation are outside these rows.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.chat reaches Muse Wonder and Muse Spark through the Responses API, the format Meta recommends. Runs stateless with store false. Meta also serves Chat Completions and Messages, which RubyLLM does not declare for this provider.","sources":["s979","s980","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality meta/muse-spark-1.3-contributor can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_meta_muse-spark-1_3-contributor_can_have_a_basic_conversation.yml","notes":"A recorded conversation through POST /v1/responses with store false."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"Server-sent events from POST /v1/responses, including function call arguments and usage on the final event.","sources":["s980"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/streaming.rb"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses meta/muse-spark-1.3-contributor supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_meta_muse-spark-1_3-contributor_supports_streaming_responses.yml","notes":"Recorded server-sent events assembled into a message."}]},"structured_output":{"v1":"missing","v2":"native","offered":true,"notes":"with_schema renders text.format with json_schema, which Meta constrains during decoding. Recursive schemas return HTTP 400, and tool search cannot be combined with a schema.","sources":["s982","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"schema_format"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat #with_schema with meta/muse-spark-1.3-contributor accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_meta_muse-spark-1_3-contributor_accepts_a_json_schema_and_returns_structured_output.yml","notes":"The request carries text.format with json_schema."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Images travel as input_image parts. The Meta dialect always sends them inline as data URIs, including remote images, because Meta refuses to fetch many remote images. Up to 50 images per request.","sources":["s983"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"render_image"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models meta/muse-spark-1.3-contributor can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_meta_muse-spark-1_3-contributor_can_understand_local_images.yml","notes":"A local image sent as input_image."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta sends remote images inline, because Meta refuses to fetch many of them","notes":"Remote images travel as data URIs."}]},"pdf_input":{"v1":"missing","v2":"native","offered":true,"notes":"PDFs travel as input_file parts: inline as file_data, or by file_url when the attachment is a URL. Meta reads text from the first 100 pages and images from the first 50.","sources":["s984"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"document_urls?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/media.rb","method":"render_document"}],"verification":[{"status":"passed","recorded_on":"2026-10-25","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat document URL models meta/muse-spark-1.3-contributor reads PDFs the provider fetches from their URL","cassette":"spec/fixtures/vcr_cassettes/chat_document_url_models_meta_muse-spark-1_3-contributor_reads_pdfs_the_provider_fetches_from_their_url.yml","notes":"A PDF URL sent as input_file with file_url, fetched by Meta."}]},"audio_input":{"v1":"missing","v2":"native","offered":true,"notes":"Audio attachments render as input_audio parts with base64 data and a format, for MP3 and WAV, on Muse Spark 1.2 and Muse Wonder. Meta calls audio on Muse Spark 1.3 not fully supported. Uploaded audio travels as input_file by file_id; a live upload of a WAV was answered on 2026-11-10.","sources":["s985","s984","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"supported_attachment?, render_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"muse-spark-1.2","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models meta/muse-spark-1.2 can understand audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_meta_muse-spark-1_2_can_understand_audio.yml","notes":"WAV sent as input_audio."},{"status":"passed","recorded_on":"2026-11-10","model":"muse-spark-1.2","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat audio models meta/muse-spark-1.2 can understand MP3 audio","cassette":"spec/fixtures/vcr_cassettes/chat_audio_models_meta_muse-spark-1_2_can_understand_mp3_audio.yml","notes":"MP3 sent as input_audio with format mp3."}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"Video attachments render as input_video parts with a data URL. Remote videos are downloaded and sent inline, because many hosts refuse Meta's fetcher. Uploaded videos travel as input_file by file_id; a live upload of an MP4 was answered on 2026-11-10. Meta documents MP4 only.","sources":["s985","s984","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"supported_attachment?, render_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"muse-spark-1.2","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models meta/muse-spark-1.2 can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_meta_muse-spark-1_2_can_understand_local_videos.yml","notes":"Request carries a data URL in input_video."},{"status":"passed","recorded_on":"2026-11-10","model":"muse-spark-1.2","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models meta/muse-spark-1.2 can understand remote videos without extension","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_meta_muse-spark-1_2_can_understand_remote_videos_without_extension.yml","notes":"The remote video is fetched and sent inline."}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"Responses return reasoning items whose encrypted content RubyLLM keeps as the thinking signature and replays across turns. Meta never returns raw reasoning as text. A readable summary comes back only when requested, with with_thinking display: :summarized, and may be empty.","sources":["s987","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_reasoning_item, parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking meta/muse-spark-1.3-contributor returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_meta_muse-spark-1_3-contributor_returns_thinking_when_available.yml","notes":"The response carries a reasoning item with encrypted content."},{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking meta/muse-spark-1.3-contributor preserves thinking signatures between turns when provided","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_meta_muse-spark-1_3-contributor_preserves_thinking_signatures_between_turns_when_provided.yml","notes":"The encrypted reasoning is replayed on the next turn."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking(effort:) renders reasoning.effort, from minimal to max. Meta rejects none, so reasoning cannot be turned off, and max is not available on muse-spark-1.3-contributor.","sources":["s987"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta chat sets the reasoning effort","notes":"with_thinking(effort: :max) renders reasoning.effort."}]},"citations":{"v1":"missing","v2":"native","offered":true,"notes":"Web search answers carry url_citation annotations with offsets, which RubyLLM parses into Message citations.","sources":["s988"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_citations, normalize_annotation"}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM tools render as flat function tools, with function_call and function_call_output items replayed statelessly. Generated arguments are not guaranteed to match the schema.","sources":["s989"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_function_call_items"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling meta/muse-spark-1.3-contributor can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_meta_muse-spark-1_3-contributor_can_use_tools.yml","notes":"A function call and its result round trip."}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"The tool_choice parameter accepts only auto, the default. Meta returns HTTP 400 for none, required, and named functions, so with_tool_options(choice:) reaches the API and Meta rejects it.","sources":["s989","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"build_tool_choice"}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"The model can return several function_call items in one response, and RubyLLM runs each one.","sources":["s989"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling meta/muse-spark-1.3-contributor can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_meta_muse-spark-1_3-contributor_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Several function_call items in one response."}]},"parallel_tool_control":{"v1":"missing","v2":"native","offered":true,"notes":"with_tool_options(calls: :one) renders parallel_tool_calls false. The flag applies only with function tools.","sources":["s989"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) renders the web_search tool and asks for its sources. Options such as user_location and search_context_size pass through the alias. Billed per 1,000 queries on top of tokens. Muse Image rejects this tool and searches by itself.","sources":["s988","s990","s991"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"SERVER_TOOL_ALIASES, web_search_sources?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_provider_tools"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with meta/muse-spark-1.3-contributor searches, returns the sources, and replays them","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_meta_muse-spark-1_3-contributor_searches_returns_the_sources_and_replays_them.yml","notes":"A web_search tool, its sources, and a replayed second turn."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta chat grounds answers in web search, and refuses other provider tools","notes":"The web_search alias renders; code_execution raises."}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool union lists function, computer, web search, and tool search. Opening a page is an internal step of web_search, not a separate tool.","sources":["s992","s993","s986","s989"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"Tool calling names web_search and tool_search as the built-in tools. Code execution means a function tool the application runs in its own sandbox. Muse Image runs code internally, which is not a tool for callers.","sources":["s992","s993","s989","s994"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool union has no file search tool. Files reach the model only as inputs.","sources":["s992","s993","s986","s989"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool union has no remote MCP tool. Meta describes namespace tools for names that come from MCP clients, which the application runs itself.","sources":["s992","s993","s986","s989"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"Prefix caching is automatic on every request. RubyLLM reads cached_tokens into the cache_read bucket, billed at the cached input rate.","sources":["s995","s990"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_usage"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta chat sends a cache key and retention without breakpoints","notes":"with_caching renders prompt_cache_key and prompt_cache_retention, and no breakpoint."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Meta documents no breakpoints: the prefix is matched from the start of the prompt. The dialect leaves message content unmarked.","sources":["s995"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"inject_cache_breakpoint"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no cache resources. Caching is automatic prefix reuse with an optional key.","sources":["s992","s993","s995"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"compaction":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Meta's Responses page says truncation auto is rejected with HTTP 400 and that the API does not trim history; applications compact it themselves. On computer requests, truncation auto trims old screenshots where it is enabled for the account. A live check on 2026-11-10 split by model: muse-wonder-1.0 rejected auto, muse-spark-1.3 accepted it, which does not show anything is trimmed. with_provider_options(truncation: \"auto\") forwards it. with_compaction sends nothing on this provider, and RubyLLM has no computer tool to pair it with.","sources":["s994","s980","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"apply_compaction, COMPACTION_PROVIDERS"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"apply_compaction"}],"verification":[{"status":"unavailable","checked_on":"2026-11-10","model":"muse-wonder-1.0, muse-spark-1.3","notes":"Live: truncation auto answered 400 \"only disabled is supported\" on muse-wonder-1.0 and was accepted on muse-spark-1.3 (status incomplete only because of max_output_tokens); disabled worked on both."}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#count_tokens posts to /v1/responses/input_tokens, which applies the model tokenizer, instructions, tools, and media.","sources":["s996","s997"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"TokenCounting"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/token_counting.rb","method":"count_tokens_url, parse_count_tokens_response"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta counts input tokens with the Responses endpoint","notes":"count_tokens posts to /v1/responses/input_tokens."}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint with muse-image-1.0 posts to /v1/images/generations through operations the Responses protocol inherits, and reads the base64 image and output_format. No recording covers this path yet. The conversational Responses route is listed under the image generation tool.","sources":["s991","s998","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"images_url, render_image_payload, parse_image_responses"}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint with with: posts multipart to /v1/images/edits. Meta returns HTTP 400 for mask. No recording covers this path yet.","sources":["s991","s999"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"render_edit_payload, build_upload_part"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The model lines and inventory list no video generation. Video is an input for Muse Spark and SAM.","sources":["s992","s993","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"Meta states that Muse Voice Transcribe does not synthesize speech, and the inventory lists no speech endpoint.","sources":["s992","s993","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts settings and audio to POST /v1/asr/transcribe as request and audio multipart parts. Meta takes WAV only: mono 16-bit PCM at 16 or 24 kHz, up to 10 minutes. language: is sent as languageBias in Meta's own words, and prompt: as keywords. timestamps: :segment asks for ENDPOINTING timed turns; word timestamps are not offered.","sources":["s1000","s1001","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"parse_transcription_response"},{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"Transcription"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_can_transcribe_audio.yml","notes":"The multipart request returns the transcript."},{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 can transcribe with language hint","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_can_transcribe_with_language_hint.yml","notes":"A language hint is accepted as languageBias."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on transcribe sets Accept: text/event-stream on /v1/asr/transcribe. Partial transcripts arrive as partial chunks, then the final transcript as a delta or each completed turn as a segment. The /v1/asr/realtime WebSocket, which takes paced live PCM, is not integrated.","sources":["s1000","s1002"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"stream_transcription"},{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription/stream.rb","method":"process"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 streams a transcript with no turns","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_streams_a_transcript_with_no_turns.yml","notes":"Partial events and a final transcript arrive over Server-Sent Events."},{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 streams turns labelled with speakers","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_streams_turns_labelled_with_speakers.yml","notes":"Each completed turn streams as a labelled segment."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"speaker_names: [] sets mode DIARIZATION on both the buffered and streaming paths, and each turn is a segment with a speaker letter. Meta does not identify named speakers, so names and speaker_references are rejected.","sources":["s1000"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"render_transcription_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 labels turns with speakers when speaker names are given","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_labels_turns_with_speakers_when_speaker_names_are_given.yml","notes":"Two speakers in the same file come back as turns labelled A and B."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no moderation endpoint. Moderation on image requests is a parameter Meta rejects on muse-image-1.0.","sources":["s992","s993","s991"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no OCR or document parsing endpoint. PDF text extraction happens inside chat input and is covered under PDF input.","sources":["s992","s993","s984"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no embeddings endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no embeddings endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no reranking endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload calls POST /v1/files through the OpenAI Files protocol. purpose: is required, user_data or batch, and expires_in: maps to expires_after. Uploaded files are referenced in Responses input and by auto_upload_large_files.","sources":["s984","s1003"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/openai/files.rb","method":"render_upload_fields"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads meta uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_meta_uploads_a_pdf_through_the_files_api.yml","notes":"POST /v1/files returns the file id and byte size."},{"status":"passed","recorded_on":"2026-11-09","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads meta/muse-spark-1.3-contributor reuses an uploaded file in chat","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_meta_muse-spark-1_3-contributor_reuses_an_uploaded_file_in_chat.yml","notes":"The file id is read back through a Responses input_file block."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download calls GET /v1/files/{file_id}/content and returns a DownloadedFile with save and to_blob; UploadedFile.find reads the metadata.","sources":["s1003"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"download"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find"},{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads meta downloads an uploaded file","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_meta_downloads_an_uploaded_file.yml","notes":"The downloaded bytes match the uploaded PDF."}]},"chat_batches":{"v1":"na","v2":"unknown","offered":null,"notes":"GET /v1/batches returned an empty list, against 404 for a nonsense path, and POST /v1/batches returned 404: batch API not enabled, so a batch route exists behind a flag this key does not have. The API reference and documentation index list no batch endpoint, and the Files API accepts purpose \"batch\". Whether a batch can run, and of what, is unverified. RubyLLM declares no batch operation for Meta.","sources":["s1004","s984","s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"ensure_batches_supported!"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/batches returned 200 with an empty list. POST /v1/batches returned 404: batch API not enabled."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no embeddings endpoint to batch.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no advisor tool.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row covers the Perplexity Agent API, which Meta does not serve.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"Skills belong to Muse Code, a separate coding agent product. The API inventory has no skills resource.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no research job endpoint. Background responses are a separate row.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"background_responses":{"v1":"missing","v2":"partial","offered":true,"notes":"Meta documents background true with polling, stream retrieval, and cancel. with_provider_options(background: true, store: true) can start one, but RubyLLM reads the queued acknowledgement as a finished turn and has no retrieval, polling, or cancel lifecycle.","sources":["s980","s997","s990"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload, parse_completion_body"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool union has no browser tool. browser.search, browser.open, and browser.find are names reserved inside web_search.","sources":["s992","s993","s988","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no classification endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"computer_use":{"v1":"missing","v2":"partial","offered":true,"notes":"Meta documents a client-executed computer tool on Responses for Muse Wonder and Muse Spark. with_provider_tools({ type: \"computer\" }) forwards the raw tool, and computer_call items come back as ServerToolCall objects. RubyLLM has no way to dispatch them or send computer_call_output items with screenshots and safety acknowledgements.","sources":["s994","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"normalize_positional"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"RESERVED_NAMESPACES"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no controls for editing context on requests. Screenshot truncation on computer requests is recorded under Context compaction, and old reasoning is the application's to drop.","sources":["s992","s993","s980","s994"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no embeddings endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe uses the dedicated POST /v1/asr/transcribe endpoint, with JSON and event-stream responses. This row overlaps Transcription.","sources":["s1000","s1001"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/meta/transcription.rb","method":"transcription_url"},{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"Transcription"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"muse-voice-transcribe-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality meta/muse-voice-transcribe-1.0 can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_meta_muse-voice-transcribe-1_0_can_transcribe_audio.yml","notes":"The multipart request to asr/transcribe returns the transcript."}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no dubbing endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"Requests with service_tier flex or priority are accepted and served as default: the response reports service_tier default for both. The schema lists auto, default, flex, scale and priority, and the guides and pricing page describe no tiers. with_provider_options(service_tier:) forwards the field.","sources":["s986","s990","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"muse-spark-1.1","notes":"POST /responses with service_tier flex and priority returned service_tier default. A nonsense value returned 400."}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"Meta documents listing, retrieving, and deleting files, plus file expiry. RubyLLM uploads, finds, and downloads Meta files and sets expiry through expires_in:, and has no public way to list or delete them.","sources":["s984","s1003"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/protocols/files.rb","method":"find"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no file search stores.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The Chat Completions schemas and the inventory document no suffix parameter or completions endpoint.","sources":["s992","s993","s979","s1005"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The Files API rejects the fine-tune purpose and the inventory lists no fine-tuning jobs. Post-training is for the open-weight models on your own GPUs or a hosted service, outside this API.","sources":["s992","s993","s1004","s1006"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no maps tool.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no guardrail resources. safety_identifier, which RubyLLM sends through with_end_user, attributes a request and configures nothing.","sources":["s992","s993","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"Meta stores responses by default and continues them with previous_response_id. RubyLLM sends store false and replays the local transcript. Meta states conversation resources are not supported on token counting, and the inventory lists no conversation endpoints.","sources":["s980","s996"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"render_payload"}]},"image_batches":{"v1":"na","v2":"unknown","offered":null,"notes":"GET /v1/batches returned an empty list, against 404 for a nonsense path, and POST /v1/batches returned 404: batch API not enabled, so a batch route exists behind a flag this key does not have. The API reference and documentation index list no batch endpoint, and the Files API accepts purpose \"batch\". Whether a batch can run, and of what, is unverified. RubyLLM declares no batch operation for Meta.","sources":["s1004","s984","s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"ensure_batches_supported!"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/batches returned 200 with an empty list. POST /v1/batches returned 404: batch API not enabled."}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists Responses, Chat Completions, and Messages. It has no Interactions API.","sources":["s992","s993","s979"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"json_mode":{"v1":"missing","v2":"passthrough","offered":true,"notes":"The schemas describe text.format type json_object as the older JSON mode, active when a message asks for JSON. with_provider_options(text: { format: { type: \"json_object\" } }) forwards it.","sources":["s986"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row covers Amazon Bedrock Mantle, which Meta does not serve.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The Responses reference lists create, retrieve, delete, cancel, and count input tokens, with no compact endpoint. Meta tells applications to compact history themselves.","sources":["s992","s993","s997","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no music endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"unknown","offered":null,"notes":"GET /v1/batches returned an empty list, against 404 for a nonsense path, and POST /v1/batches returned 404: batch API not enabled, so a batch route exists behind a flag this key does not have. The API reference and documentation index list no batch endpoint, and the Files API accepts purpose \"batch\". Whether a batch can run, and of what, is unverified. RubyLLM declares no batch operation for Meta.","sources":["s1004","s984","s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"ensure_batches_supported!"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/batches returned 200 with an empty list. POST /v1/batches returned 404: batch API not enabled."}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no video generation.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The API serves Meta models only. Its three request formats reach the same models.","sources":["s992","s993","s979","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The Chat Completions schemas and the inventory document no prefix parameter.","sources":["s992","s993","s979","s1005"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no tool that lets model-written code call other tools.","sources":["s992","s993","s989"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_caching(key:, ttl:) renders prompt_cache_key and prompt_cache_retention, in_memory or 24h. Retention is a hint and applies on Responses. Other options raise ArgumentError.","sources":["s995","s980"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"prompt_cache_params, PROMPT_CACHE_OPTIONS"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/meta_spec.rb","example":"RubyLLM::Providers::Meta chat sends a cache key and retention without breakpoints","notes":"with_caching renders prompt_cache_key and prompt_cache_retention, and no breakpoint."}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no server-side model fallback.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The /v1/asr/realtime WebSocket streams transcription only, recorded under Streaming transcription. Meta states it provides no speech-to-speech conversation API.","sources":["s992","s993","s981","s1002"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no whole-response cache. Prompt caching is a separate row.","sources":["s992","s993","s995"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"Responses is the only protocol the Meta provider declares. RubyLLM runs it stateless with encrypted reasoning replay, the path Meta recommends.","sources":["s980","s997","s979"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"Responses"}],"verification":[{"status":"passed","recorded_on":"2026-10-23","model":"muse-spark-1.3-contributor","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality meta/muse-spark-1.3-contributor can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_meta_muse-spark-1_3-contributor_can_have_a_basic_conversation.yml","notes":"A recorded conversation through POST /v1/responses with store false."}]},"responses_batches":{"v1":"na","v2":"unknown","offered":null,"notes":"GET /v1/batches returned an empty list, against 404 for a nonsense path, and POST /v1/batches returned 404: batch API not enabled, so a batch route exists behind a flag this key does not have. The API reference and documentation index list no batch endpoint, and the Files API accepts purpose \"batch\". Whether a batch can run, and of what, is unverified. RubyLLM declares no batch operation for Meta.","sources":["s1004","s984","s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"ensure_batches_supported!"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","notes":"GET /v1/batches returned 200 with an empty list. POST /v1/batches returned 404: batch API not enabled."}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no standalone search endpoint. Search runs through the web_search tool.","sources":["s992","s993","s988"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool union has no apply patch tool.","sources":["s992","s993","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Muse Image takes an image_generation tool on Responses, and rejects every other tool. Meta declares no alias for it, so with_provider_tools(:image_generation) raises. with_provider_tools({ type: \"image_generation\" }) forwards the raw tool, and image_generation_call items come back as ServerToolCall objects.","sources":["s991"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta/responses.rb","method":"SERVER_TOOL_ALIASES"},{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"normalize_positional"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"server_tool_search":{"v1":"missing","v2":"native","offered":true,"notes":"Deferred RubyLLM tools render with defer_loading and the hosted tool_search tool, and namespaces render as namespace tools. Meta documents this on Responses for any model with tool calling. No Meta recording covers it yet. Client-executed tool search is not implemented.","sources":["s1007","s986"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/responses.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/tools.rb","method":"apply_tool_search, render_tools"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no X search tool.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"missing","v2":"partial","offered":true,"notes":"Meta documents the client-executed computer tool and no shell tool for callers. The raw computer tool reaches the API, but RubyLLM has no computer_call_output lifecycle. Muse Image runs code internally, which is not a tool for callers.","sources":["s994","s986","s991"],"code":[{"version":"v2","path":"lib/ruby_llm/tools/provider_tools.rb","method":"normalize_positional"},{"version":"v2","path":"lib/ruby_llm/protocols/responses/chat.rb","method":"parse_server_tool_items"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no sound effects endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"Meta offers no speech synthesis.","sources":["s992","s993","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no task budget. max_tool_calls caps built-in tool calls only.","sources":["s992","s993","s989"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no text analysis endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no vector stores.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no video generation to batch.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no video editing endpoint. Segmentation with SAM returns geometry, not video.","sources":["s992","s993","s981"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no video generation.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no voice endpoints beyond transcription.","sources":["s992","s993","s1000"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no voice endpoints beyond transcription.","sources":["s992","s993","s1000"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory lists no standalone web fetch endpoint.","sources":["s992","s993"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API documents server-sent events only. The one WebSocket in the inventory carries transcription.","sources":["s992","s993","s980","s1002"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/meta.rb","method":"protocol"}]}},"gaps":[{"notes":"Streaming transcription: the /v1/asr/realtime WebSocket is a separate API and is not integrated. RubyLLM streams recorded files through /v1/asr/transcribe.","sources":["s1000","s1001"]},{"notes":"File listing and deletion: Meta documents both, and RubyLLM has no public way to list or delete uploaded files.","sources":["s984","s1003"]},{"notes":"Background Responses jobs: raw options can start a job, with no retrieval, polling, or cancel lifecycle.","sources":["s980"]},{"notes":"Computer use, Shell and computer tools: the raw computer tool reaches the API, with no way to return computer_call_output items.","sources":["s994"]},{"notes":"Image generation, Image editing: the operations reach the documented endpoints, and no recording covers them yet.","sources":["s991"]},{"notes":"Image generation tool: raw tool only, because Meta declares no image_generation alias.","sources":["s991"]},{"notes":"Context compaction: truncation auto works only on computer requests with limited access, and only through raw options.","sources":["s994"]},{"notes":"Chat batches, Image batches, Responses batches, Other batch operations: POST /v1/batches returned batch API not enabled, and no batch endpoint is documented, so availability is unverified.","sources":["s1004","s984"]}]},{"id":"minimax","name":"MiniMax","notes":"This covers the MiniMax API at platform.minimax.io: text, video, speech, image, music and file endpoints. RubyLLM connects to the Anthropic-compatible Messages endpoint for chat, to Chat Completion v2 only for with_schema on MiniMax-Text-01, to Responses only for web search on chats without a PDF, and to the root endpoints for images, video, speech, transcription and files. The OpenAI-compatible endpoint has no route.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.chat reaches POST /anthropic/v1/messages, the route MiniMax recommends, for MiniMax-M3.1-Flash and MiniMax-M3. The same text models also have OpenAI-compatible and Responses routes; RubyLLM uses Responses only for web search. The endpoint ignores temperature, top_p, top_k and stop_sequences.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol, api_base"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"completion_url, render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality minimax/MiniMax-M3.1-Flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_minimax_minimax-m3_1-flash_can_have_a_basic_conversation.yml","notes":"POST /anthropic/v1/messages returns a text block and usage."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"A block on ask sends stream: true on the Messages endpoint and parses its server-sent events, including thinking and signature deltas.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/streaming.rb","method":"stream_url"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses minimax/MiniMax-M3.1-Flash supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_minimax_minimax-m3_1-flash_supports_streaming_responses.yml","notes":"The stream returns Anthropic-style message and content block events."}]},"structured_output":{"v1":"na","v2":"native","offered":true,"notes":"with_schema on MiniMax-Text-01 goes to /v1/text/chatcompletion_v2 with response_format json_schema, and RubyLLM removes the object MiniMax wraps the answer in, named after the schema, so message.parsed returns the schema's object. The model is not in the registry and needs assume_model_exists: true. It cannot stream a schema, because the wrapper arrives first, and that raises ArgumentError. Every other MiniMax model ignores all structured output fields, so with_schema raises ArgumentError on them instead of returning plain text.","sources":["s1009","s1010","s1011","s1012"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/chat_completions.rb","method":"completion_url, render_payload, parse_completion_body"},{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-Text-01","spec":"spec/ruby_llm/providers/minimax/chat_completions_spec.rb","example":"RubyLLM::Providers::MiniMax::ChatCompletions returns the object the schema describes, without the name MiniMax wraps it in","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_chatcompletions_returns_the_object_the_schema_describes_without_the_name_minimax_wraps_it_in.yml","notes":"The response nests the answer under the schema name, and message.parsed returns the schema's object."},{"status":"passed","checked_on":"2026-11-09","model":"MiniMax-Text-01","notes":"Four requests with a json_schema enum returned that enum value nested under the schema name, as JSON, for \"Name a fruit.\". Without response_format the reply was a sentence."},{"status":"unavailable","checked_on":"2026-11-09","model":"MiniMax-M3.1-Flash","notes":"Every route replied \"Apple.\" to \"Name a fruit.\"; Messages also accepted a nonsense format type. RubyLLM with_schema returned \"Apple.\"."},{"status":"unavailable","checked_on":"2026-11-09","model":"MiniMax-M3","notes":"output_config.format and response_format on Messages, and json_schema on /v1/chat/completions, all replied \"Apple.\" to \"Name a fruit.\", as did M3.1-Flash. A forced tool with an enum returned {\"fruit\" => \"pear\"} from the enum on both models."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax options MiniMax cannot carry raises for with_schema before sending, because its models ignore the schema","notes":"with_schema raises ArgumentError and no request goes out."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Image attachments render as image blocks by URL or base64. MiniMax documents JPEG, PNG, GIF and WEBP for MiniMax-M3.1-Flash and MiniMax-M3. The optional detail setting has no RubyLLM keyword.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"render_image, supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models minimax/MiniMax-M3.1-Flash can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_minimax_minimax-m3_1-flash_can_understand_local_images.yml","notes":"A base64 image block returns a description of the image."}]},"pdf_input":{"v1":"na","v2":"native","offered":true,"notes":"RubyLLM renders PDF attachments as document blocks on this protocol, and the endpoint reads them. The compatibility page does not mention document blocks or PDFs.","sources":["s1008","s1009","s1013"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/media.rb","method":"render_pdf, supported_attachment?"}],"verification":[{"status":"passed","checked_on":"2026-11-09","model":"MiniMax-M3.1-Flash","notes":"Through ask(with: pdf), a one-page PDF returned its passcode in three raw requests and once through RubyLLM. Without the document the reply was NO DOCUMENT."}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The Messages, Chat Completions and Responses references list text, image and video input for the M3 family. They document no audio content block. Audio is an input only to the separate transcription and video generation operations.","sources":["s1008","s1014","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"MiniMax-M3.1-Flash and MiniMax-M3 take video attachments as video blocks. Local and remote videos travel as base64, because many hosts refuse MiniMax's fetcher, and uploaded videos as mm_file references; a live upload with purpose video_understanding was answered on 2026-11-10. MP4, AVI, MOV and MKV are documented; MP4 is recorded. Frame sampling options stay with the provider's defaults.","sources":["s1008","s1009","s1013"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"supported_attachment?, render_content, video_source"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models minimax/MiniMax-M3.1-Flash can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_minimax_minimax-m3_1-flash_can_understand_local_videos.yml","notes":"Base64 video block."},{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models minimax/MiniMax-M3.1-Flash can understand remote videos without extension","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_minimax_minimax-m3_1-flash_can_understand_remote_videos_without_extension.yml","notes":"The remote video is fetched and sent inline."}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"Thinking blocks and deltas parse into Message#thinking with their signature, and stored blocks replay on later turns as MiniMax requires. MiniMax-M3.1-Flash always thinks; MiniMax-M3 thinks only on request.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"parse_thinking_content, parse_thinking_signature, render_message"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking minimax/MiniMax-M3.1-Flash returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_minimax_minimax-m3_1-flash_returns_thinking_when_available.yml","notes":"The response carries a signed thinking block before the text."},{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking minimax/MiniMax-M3.1-Flash streams thinking content when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_minimax_minimax-m3_1-flash_streams_thinking_content_when_available.yml","notes":"The stream carries thinking_delta and signature_delta events."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking maps to thinking.type adaptive or disabled and to output_config.effort. Effort tunes MiniMax-M3.1-Flash, from low to max. MiniMax-M3 has an on and off toggle and ignores effort. with_thinking(false) raises locally on MiniMax-M3.1-Flash and the M2 family, whose API also cannot turn thinking off.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"add_thinking_fields, build_thinking_payload"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_thinking"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax chat asks for adaptive thinking with an effort","notes":"with_thinking(effort: :low) sends adaptive thinking and the effort."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax chat turns thinking off on models with a toggle","notes":"with_thinking(false) on MiniMax-M3 sends disabled."}]},"citations":{"v1":"na","v2":"native","offered":true,"notes":"Web search routes to the Responses endpoint, which returns url_citation annotations that RubyLLM parses into typed Citation objects with the quoted snippet and no offsets, since MiniMax sends both as 0, in sync and streamed replies. A chat that holds a PDF, sets a thinking budget or turns thinking off stays on Messages, where citations do not exist; a search run there still reports its pages through Message#web_sources.","sources":["s1008","s1009","s1015","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol_for, web_search_route?"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/responses.rb","method":"parse_annotations, offset_citations, parse_web_sources"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"complete, render"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax web search citations returns the pages a web search cites","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_web_search_citations_returns_the_pages_a_web_search_cites.yml","notes":"A web search on Responses returned url_citation annotations as Citation objects and web sources."},{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax web search citations streams the pages a web search cites","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_web_search_citations_streams_the_pages_a_web_search_cites.yml","notes":"The same citations on a stream."}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"Ruby tools render as name, description and input_schema, and tool_use blocks round-trip with their tool results. MiniMax asks for the full assistant content in later turns, including thinking, and RubyLLM replays it.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"render_tools, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling minimax/MiniMax-M3.1-Flash can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_minimax_minimax-m3_1-flash_can_use_tools.yml","notes":"A tool_use block runs a Ruby tool and its result returns a final answer."}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"auto, the default, is the only tool_choice value MiniMax documents on any of its text routes, so there is no choice to control. with_tool_options(choice:) sends other values as given, and the documentation does not say how the endpoint answers them.","sources":["s1008","s1009","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_tool_options"}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"Several tool_use blocks in one response parse into separate tool calls. The MiniMax documentation does not describe this on the Messages route; the recording shows it, and the Responses reference reports parallel_tool_calls as always true.","sources":["s1009","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling minimax/MiniMax-M3.1-Flash can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_minimax_minimax-m3_1-flash_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"One response returns three dice_roll tool_use blocks."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"disable_parallel_tool_use is accepted and ignored: four parallel calls came back in every run on MiniMax-M3.1-Flash with it set to true, and also on Responses with parallel_tool_calls false, as documented. Unknown fields are also accepted. Nothing on Messages limits the calls, so with_tool_options(calls: :one) raises ArgumentError instead of sending a field that does nothing.","sources":["s1009","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"render_payload"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"MiniMax-M3.1-Flash","notes":"Four requests with tool_choice { type: auto, disable_parallel_tool_use: true } returned four tool calls, the same as the default. A top-level parallel_tool_calls: false on Messages did the same."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax options MiniMax cannot carry raises for calls: :one before sending, because the API ignores it","notes":"calls: :one raises ArgumentError and no request goes out."}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) runs on the Responses endpoint, which cites its pages. It stays on Messages as web_search_20250305 when the chat holds a PDF, which Responses ignores, or sets a thinking budget or turns thinking off, or uses calls: :one; pause turns there continue automatically. Either way the search steps appear as server tool calls and Message#web_sources lists the pages. MiniMax marks server tools as Beta.","sources":["s1015","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol_for, web_search_route?"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/responses.rb","method":"SERVER_TOOL_ALIASES, parse_web_sources"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"SERVER_TOOL_ALIASES, server_tool_aliases"},{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"MAX_PAUSE_TURN_CONTINUATIONS, parse_web_sources"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax web search citations returns the pages a web search cites","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_web_search_citations_returns_the_pages_a_web_search_cites.yml","notes":"A live search on Responses with its pages."},{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax web search citations keeps reading a PDF attached after a cited web search turn","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_web_search_citations_keeps_reading_a_pdf_attached_after_a_cited_web_search_turn.yml","notes":"A later PDF moved the chat to Messages and was read."}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool. RubyLLM raises UnsupportedServerToolError for this alias on MiniMax.","sources":["s1015","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"SERVER_TOOL_ALIASES"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool, and the Messages route ignores container. RubyLLM raises UnsupportedServerToolError for this alias on MiniMax.","sources":["s1015","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"SERVER_TOOL_ALIASES"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax rejects the other provider tools, which the endpoint does not run","notes":"with_provider_tools(:code_execution) raises at request time."}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool and offers no hosted file search.","sources":["s1015","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"SERVER_TOOL_ALIASES"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The Messages route documents mcp_servers as ignored, and web_search is the only server tool. MiniMax publishes MCP servers for client applications to run, which is not a hosted MCP tool.","sources":["s1015","s1008","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"SERVER_TOOL_ALIASES"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"MiniMax caches automatically from 512 input tokens by prefix, with no request option. RubyLLM reports usage.cache_read_input_tokens as cache_read, and input_tokens holds new input only, matching how MiniMax bills.","sources":["s1017","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_message"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax chat parses thinking with its signature, and counts cached input","notes":"cache_read_input_tokens becomes tokens.cache_read."},{"status":"passed","recorded_on":"2026-10-20","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality minimax/MiniMax-M3.1-Flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_minimax_minimax-m3_1-flash_can_have_a_basic_conversation.yml","notes":"The response reports 197 cache read tokens on an automatic cache hit."}]},"cache_boundaries":{"v1":"missing","v2":"na","offered":false,"notes":"Not served. The Messages schema states that MiniMax-M3 and MiniMax-M3.1-Flash do not support cache_control, and every other model that took it (the M2 family) is on the Retired list and answers not_found_error 2069. Caching is automatic: with cache_control on a new prefix M3.1 Flash reports no cache write, and a repeat of the same prefix reads the cache either way. with_caching and cache_until_here still render the markers, which these models ignore.","sources":["s1009","s1012","s1017"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"build_system_content, render_payload"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_caching, cache_until_here"}],"verification":[{"status":"unavailable","checked_on":"2026-11-10","model":"MiniMax-M3.1-Flash","notes":"Live: MiniMax-M2, M2.1 and M2.7 answered not_found_error 2069 \"no longer available\". On M3.1 Flash a fresh prefix with cache_control returned cache_creation_input_tokens 0 and cache_read_input_tokens 0, and the prefix repeated returned a cache read through automatic caching."}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The Prompt Caching page describes automatic caching only. The API inventory documents no cache resources to create or manage.","sources":["s1017","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"The Messages route documents context_management as ignored, and the other text routes document no compaction or sliding window control. The Responses truncation field is documented as ignored too. with_compaction still sends context_management, which has no effect here.","sources":["s1008","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"apply_compaction"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_compaction"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"chat.count_tokens posts to /anthropic/v1/messages/count_tokens, which MiniMax documents for MiniMax-M3.1-Flash and MiniMax-M3. MiniMax also has a separate Responses estimate at /v1/responses/input_tokens. No recording covers this route.","sources":["s1008","s1018"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic/chat.rb","method":"count_tokens_url, render_count_tokens_payload, parse_count_tokens_response"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"count_tokens"}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint posts to POST /v1/image_generation with the image-01 family and returns base64 data by default; provider_options: { response_format: \"url\" } returns 24 hour links instead. size: takes WIDTHxHEIGHT or an aspect ratio, and count: up to 9 images.","sources":["s1019","s1020","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/images.rb","method":"render_image_payload, parse_image_responses"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"image-01","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality minimax/image-01 can paint images","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_minimax_image-01_can_paint_images.yml","notes":"The recording returns a JPEG in base64."},{"status":"passed","recorded_on":"2026-11-09","model":"image-01","spec":"spec/ruby_llm/image_spec.rb","example":"RubyLLM::Image basic functionality minimax/image-01 supports custom sizes","cassette":"spec/fixtures/vcr_cassettes/image_basic_functionality_minimax_image-01_supports_custom_sizes.yml","notes":"A 1792x1024 request sent width and height."}]},"image_editing":{"v1":"missing","v2":"native","offered":true,"notes":"with: on paint sends subject_reference on POST /v1/image_generation, which keeps a person from a URL or image file in a new scene. MiniMax documents only the character type and no mask, so mask: raises UnsupportedAttachmentError.","sources":["s1019","s1021"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/images.rb","method":"subject_references, validate_paint_inputs!"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"image-01","spec":"spec/ruby_llm/providers/minimax_media_spec.rb","example":"RubyLLM::Providers::MiniMax live keeps a person from a reference image in a new scene","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_live_keeps_a_person_from_a_reference_image_in_a_new_scene.yml","notes":"A portrait URL produced a new JPEG."}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate creates a task on POST /v2/video_generation and polls GET /v2/query/video_generation/{id} for MiniMax-H3 models. It defaults resolution 768P, duration 5, and ratio 16:9 for text-only requests, which the endpoint requires; provider_options: changes them. One image is the first frame and other attachments are references. Legacy Hailuo task endpoints have no route.","sources":["s1022","s1023","s1016","s1012"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/videos.rb","method":"render_video_payload, parse_video_job_status, download_video"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"MiniMax-H3-Lite","spec":"spec/ruby_llm/video_spec.rb","example":"RubyLLM::Video basic functionality minimax/MiniMax-H3-Lite can animate videos","cassette":"spec/fixtures/vcr_cassettes/video_basic_functionality_minimax_minimax-h3-lite_can_animate_videos.yml","notes":"A 4 second 480P task succeeded and the MP4 downloaded."}]},"speech":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.speak posts to POST /v1/t2a_v2 with the speech model family, decoding the hex audio from the JSON body. voice: maps to voice_setting.voice_id and format: to audio_setting.format. Streaming and the asynchronous long text route have no RubyLLM path.","sources":["s1024","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/speech.rb","method":"render_speech_payload, parse_speech_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"speech-2.8-turbo","spec":"spec/ruby_llm/speech_spec.rb","example":"RubyLLM::Speech basic functionality minimax/speech-2.8-turbo can speak","cassette":"spec/fixtures/vcr_cassettes/speech_basic_functionality_minimax_speech-2_8-turbo_can_speak.yml","notes":"The recording decodes to MP3 audio."}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts multipart to POST /v1/speech_to_text with asr-1.0, up to 500 seconds and 50 MB. language: goes in the language header. format: selects json, verbose_json, srt or vtt, and timestamps: asks for word or segment units. prompt and temperature have no field and raise ArgumentError.","sources":["s1025","s1026"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/transcription.rb","method":"transcribe, render_transcription_payload, parse_transcription_response"},{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"resolve_protocol, media_base"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"asr-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality minimax/asr-1.0 can transcribe audio","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_minimax_asr-1_0_can_transcribe_audio.yml","notes":"The recording returns the transcript and duration."},{"status":"passed","recorded_on":"2026-11-09","model":"asr-1.0","spec":"spec/ruby_llm/transcription_spec.rb","example":"RubyLLM::Transcription basic functionality minimax/asr-1.0 can transcribe with language hint","cassette":"spec/fixtures/vcr_cassettes/transcription_basic_functionality_minimax_asr-1_0_can_transcribe_with_language_hint.yml","notes":"The language header is accepted and the English transcript returned."}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on RubyLLM.transcribe sends stream=true and reads the server-sent deltas, with the final event carrying the duration. Only the json format streams; audio input cannot be streamed.","sources":["s1025","s1026"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/transcription.rb","method":"stream_transcription"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"asr-1.0","spec":"spec/ruby_llm/providers/minimax_media_spec.rb","example":"RubyLLM::Providers::MiniMax live streams a transcription","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_live_streams_a_transcription.yml","notes":"Deltas join to the transcript, and the last chunk is done."}]},"diarization":{"v1":"missing","v2":"native","offered":true,"notes":"speaker_names: [] asks for verbose_json, which returns n_speakers and a speaker tag such as S1 on each segment in Transcription#segments. MiniMax cannot name speakers, so a non-empty speaker_names or speaker_references raises ArgumentError. It cannot be combined with streaming.","sources":["s1025","s1026"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/transcription.rb","method":"render_transcription_payload, rejected_options"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"asr-1.0","spec":"spec/ruby_llm/providers/minimax_media_spec.rb","example":"RubyLLM::Providers::MiniMax live labels speakers","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_live_labels_speakers.yml","notes":"The recording returns timed segments labelled S1."}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no moderation operation. Sensitive content flags on legacy text responses and the mask_sensitive_info option do not classify input on request.","sources":["s1016","s1027","s1011"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no OCR or document parsing endpoint. Reading an image through chat is not an OCR API.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no embeddings endpoint. RubyLLM raises on embed through this protocol.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no embeddings endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no reranking endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to POST /v1/files/upload and requires a purpose: from MiniMax (voice_clone, prompt_audio, t2a_async_input, video_understanding, video_generation_input). Uploaded images, video and audio are referenced by mm_file in animate; chat attachments stay inline, since the Messages dialect turns provider file references off.","sources":["s1013","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/files.rb","method":"render_upload_fields, parse_file_response"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"provider_file_references?"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"n/a","spec":"spec/ruby_llm/providers/minimax_media_spec.rb","example":"RubyLLM::Providers::MiniMax live uploads a file","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_live_uploads_a_file.yml","notes":"A t2a_async_input text file returned its file_id, filename and size."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download calls GET /v1/files/retrieve_content, which serves generated files such as long text speech output. It refused an uploaded t2a_async_input file with \"invalid file purpose\", so user uploads cannot be read back.","sources":["s1028","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/files.rb","method":"download_file_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"n/a","spec":"spec/ruby_llm/providers/minimax_media_spec.rb","example":"RubyLLM::Providers::MiniMax live downloads a generated file","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_live_downloads_a_generated_file.yml","notes":"A long text speech result downloaded as a tar archive."}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint for chat. Asynchronous long text speech is a single task, not a batch.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no embeddings or batch endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no advisor tool.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no Perplexity Agent API; that row belongs to another provider.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no agent skills, and the Messages route ignores container.","sources":["s1016","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no asynchronous research operation.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Create Response reference documents no background parameter. Asynchronous video and long text speech tasks are separate operations.","sources":["s1010","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no browser use tool.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no classification endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool, and tool definitions take a name and JSON schema only.","sources":["s1015","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The Messages route documents context_management as ignored.","sources":["s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no contextual embeddings.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe reaches POST /v1/speech_to_text, the dedicated endpoint with JSON, verbose JSON, SRT and VTT output. The format: option selects it.","sources":["s1025","s1026"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/minimax/transcription.rb","method":"transcribe, transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no dubbing operation.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"fast_inference":{"v1":"missing","v2":"passthrough","offered":true,"notes":"service_tier: \"priority\" buys priority admission at 1.5 times the standard price on MiniMax-M3; MiniMax-M3.1-Flash does not support the field. with_provider_options(service_tier: \"priority\") forwards it, and RubyLLM has no shared keyword for it.","sources":["s1008","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"render"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"GET /v1/files/list and POST /v1/files/delete manage stored files. RubyLLM has no list or delete operation.","sources":["s1029","s1030","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no file search stores.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no fill-in-the-middle parameter or endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The hosted API documents no fine-tuning jobs. The LoRA recipe in the self-hosting guide trains open weights outside this API.","sources":["s1016","s1027","s1031"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no maps grounding.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no guardrail configuration.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The Create Response reference states that previous_response_id, conversation and store are not supported in requests and are ignored.","sources":["s1010","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no image batches.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no Interactions API.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"No text route documents a JSON object mode. The Text Generation API accepts only json_schema, for MiniMax-Text-01 alone, and the Responses format enum lists only text.","sources":["s1011","s1010"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no Bedrock Mantle protocols; that row belongs to another provider.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no compaction operation.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"music_generation":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/music_generation is limited access: it closed to new users on August 20, 2026, and existing paying users keep it. The model list also marks the music models as retired, without a separate date for those users. RubyLLM has no music operation for MiniMax.","sources":["s1032","s1012","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint. Asynchronous long text speech and video tasks are single tasks.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"omni_video_generation":{"v1":"missing","v2":"missing","offered":true,"notes":"MiniMax-H3 takes a content array that mixes text with reference images, videos and audio on POST /v2/video_generation. RubyLLM.animate sends reference attachments in that array, but mixed references have no live check.","sources":["s1022","s1023"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax serves its own models only, so no partner model routes exist in this inventory.","sources":["s1016","s1012"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The text references document no assistant prefill or prefix parameter.","sources":["s1008","s1009","s1014"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool and no programmatic tool calling.","sources":["s1015","s1008"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"missing","v2":"missing","offered":true,"notes":"The Responses route takes prompt_cache_key as a cache routing identifier. On a web search chat routed there, with_caching(key:) sends it, with no live check; the Messages route has no equivalent for current models.","sources":["s1010","s1017"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no server-side model fallback.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no realtime conversation session. Its WebSocket endpoints synthesize speech only, which the streaming speech row covers.","sources":["s1033","s1034","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no whole response cache, only prompt caching.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"responses":{"v1":"missing","v2":"partial","offered":true,"notes":"POST /v1/responses is a documented stateless route with function tools, web search and token estimates. RubyLLM registers it as protocol :responses and routes web search chats there, except those that must stay on Messages (a PDF, a thinking budget or off switch, or calls: :one). Only web search has a live check; function tools and token estimates on this route do not.","sources":["s1010","s1018","s1015"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol, protocol_for"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/responses.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"MiniMax-M3.1-Flash","spec":"spec/ruby_llm/providers/minimax_spec.rb","example":"RubyLLM::Providers::MiniMax web search citations returns the pages a web search cites","cassette":"spec/fixtures/vcr_cassettes/providers_minimax_web_search_citations_returns_the_pages_a_web_search_cites.yml","notes":"A web search chat on this route."}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no Responses batches.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The pay-as-you-go inventory documents no search endpoint. The M Plan web_search is an MCP tool for subscribers to run in their own tools, and web_search on the text routes is a server tool.","sources":["s1016","s1035","s1015"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no apply patch tool.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no hosted image generation tool.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool. RubyLLM treats this protocol as searching deferred tools itself, so deferred tools send a tool search definition that the documentation does not list.","sources":["s1015","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/anthropic.rb","method":"tool_search?"},{"version":"v2","path":"lib/ruby_llm/providers/minimax/anthropic.rb","method":"server_tool_aliases"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no X search tool.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"MiniMax documents web_search as its only server tool, and tool definitions take a name and JSON schema only.","sources":["s1015","s1009"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no sound effect generation. The sound_effects setting on speech applies a filter to synthesized voice.","sources":["s1016","s1024"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"streaming_speech":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/t2a_v2 streams audio with stream: true, and two WebSocket endpoints stream audio, one of them from streaming text input. RubyLLM has no route for it.","sources":["s1024","s1033","s1034"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no task budget control.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no text analysis endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no vector stores.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no video batches.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no general video edit. Video regeneration only turns a MiniMax-H3 768P result into 2K, and the page states that it does not process arbitrary videos.","sources":["s1036","s1023","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The video references document no operation that extends an existing video.","sources":["s1022","s1023","s1036"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"voice_cloning":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/voice_clone creates a voice_id from uploaded audio, after individual or enterprise verification. Cloned voices are deleted after 7 days without use. RubyLLM has no voice creation operation for MiniMax.","sources":["s1037","s1038","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no speech-to-speech conversion.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no standalone web fetch endpoint.","sources":["s1027","s1016"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The Create Response reference documents HTTP only. MiniMax WebSocket endpoints serve speech synthesis, not Responses.","sources":["s1010","s1033"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/minimax.rb","method":"protocol"}]}},"gaps":[{"notes":"Tool choice: MiniMax documents only auto, so forcing or disabling a tool is not available.","sources":["s1008","s1009"]},{"notes":"Omni video generation: animate sends references in content, but mixed references have no live check.","sources":["s1019","s1021","s1022"]},{"notes":"Streaming speech generation: RubyLLM takes the synchronous T2A route only, so stream: true and the WebSocket endpoints have no route.","sources":["s1024","s1033","s1034"]},{"notes":"File listing and deletion: RubyLLM has no operation for GET /v1/files/list or POST /v1/files/delete, and chat attachments other than video stay inline rather than using mm_file.","sources":["s1013","s1028","s1029","s1030"]},{"notes":"Responses API, Prompt cache options: POST /v1/responses and its prompt_cache_key are documented. RubyLLM routes only web search chats there, where with_caching(key:) sends a cache key with no live check.","sources":["s1010"]},{"notes":"Music generation: limited to existing paying users since August 20, 2026, with no RubyLLM operation for it.","sources":["s1032","s1012"]},{"notes":"Voice cloning: MiniMax creates voices from uploaded audio, and RubyLLM has no voice creation operation for it.","sources":["s1037","s1038"]},{"notes":"Fast inference: priority admission needs service_tier through with_provider_options, on MiniMax-M3 only.","sources":["s1008"]},{"notes":"Structured output: with_schema works only on MiniMax-Text-01, on Chat Completion v2, without streaming; the M3 models ignore every schema field, so with_schema raises ArgumentError on them.","sources":["s1009","s1010"]}]},{"id":"moonshot","name":"Moonshot AI","notes":"Audited the Kimi API Platform at api.moonshot.ai: its REST inventory, OpenAPI file, guides, and changelog. RubyLLM integrates Chat Completions, Responses for web search, the Files API and the Batch API. Admin API endpoints are out of comparison.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.chat calls /v1/chat/completions on api.moonshot.ai with Kimi K3.1 Pro, K3, K2.7 Code, and K2.6. kimi-k2.6 and kimi-k2.7-code are closed to new accounts, with deprecation expected in late December 2026.","sources":["s1039","s1040"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_completion_body"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality moonshot/kimi-k3 can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_moonshot_kimi-k3_can_have_a_basic_conversation.yml","notes":"Successful Chat Completions text response."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"Chat streaming reads server-sent events, including reasoning_content deltas and usage from the final chunk. RubyLLM sends stream_options include_usage.","sources":["s1041"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses moonshot/kimi-k3 supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_moonshot_kimi-k3_supports_streaming_responses.yml","notes":"Streamed text response."}]},"structured_output":{"v1":"missing","v2":"native","offered":true,"notes":"with_schema sends response_format json_schema. Kimi describes K2.6 as unstable with complex schemas; K3 and K3.1 Pro handle nested objects, arrays, and anyOf.","sources":["s1042"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload, schema_strict"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"kimi-k3","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat #with_schema with moonshot/kimi-k3 accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_moonshot_kimi-k3_accepts_a_json_schema_and_returns_structured_output.yml","notes":"Parsed structured output on K3."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Image attachments travel as base64 data URIs in image_url parts. Kimi does not read image URLs, so RubyLLM downloads remote images and sends them inline. An image uploaded with RubyLLM.upload goes as an ms:// reference.","sources":["s1043","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_content, render_inline_image"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models moonshot/kimi-k3 can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_moonshot_kimi-k3_can_understand_local_images.yml","notes":"Local image sent as a data URI."},{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models moonshot/kimi-k3 can understand remote images without extension","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_moonshot_kimi-k3_can_understand_remote_images_without_extension.yml","notes":"Remote image downloaded and sent inline."}]},"pdf_input":{"v1":"missing","v2":"native","offered":true,"notes":"Kimi models cannot read an attached file, and Kimi documents one flow for PDF questions: extract the text through the Files API and put it in messages. RubyLLM does that for PDF, Word and spreadsheet attachments and sends the Markdown as a text part, so figures in a document are not seen. Uploads are reused across turns and stay in the account's file storage.","sources":["s1044","s1045","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/documents.rb","method":"preprocess_message, extracted_attachment"},{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"supported_attachment?"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models moonshot/kimi-k3 understands PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_moonshot_kimi-k3_understands_pdfs.yml","notes":"PDF uploaded, its text read back and sent in the message; the follow-up turn reuses the upload."},{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat pdf models moonshot/kimi-k3 handles multiple PDFs","cassette":"spec/fixtures/vcr_cassettes/chat_pdf_models_moonshot_kimi-k3_handles_multiple_pdfs.yml","notes":"Local and remote PDFs in one message."}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The Chat Completions content types are text, image_url, and video_url. The audited API inventory documents no audio input, and RubyLLM rejects audio attachments for this provider.","sources":["s1046","s1047","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"supported_attachment?"}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"K3.1 Pro, K3, K2.7 Code and K2.6 take video attachments as video_url parts on Chat Completions. Kimi reads only data URIs and its own file references, so remote videos travel inline and uploaded videos as ms:// references; a live upload was answered on 2026-11-10. Video on the Responses route, used for web search, still raises UnsupportedAttachmentError.","sources":["s1043","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"supported_attachment?, render_inline_video, inline_url"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"kimi-k2.6","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models moonshot/kimi-k2.6 can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_moonshot_kimi-k2_6_can_understand_local_videos.yml","notes":"Data URI in video_url."},{"status":"passed","recorded_on":"2026-11-10","model":"kimi-k2.6","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models moonshot/kimi-k2.6 can understand remote videos without extension","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_moonshot_kimi-k2_6_can_understand_remote_videos_without_extension.yml","notes":"The remote video is fetched and sent inline."}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"Message#thinking reads reasoning_content from complete and streamed responses, and RubyLLM sends it back on later assistant messages, as K3 and K2.7 Code require.","sources":["s1048"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_thinking"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"build_chunk"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking moonshot/kimi-k3 returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_moonshot_kimi-k3_returns_thinking_when_available.yml","notes":"reasoning_content on a complete response."},{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking moonshot/kimi-k3 streams thinking content when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_moonshot_kimi-k3_streams_thinking_content_when_available.yml","notes":"reasoning_content deltas while streaming."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking(effort:) sends reasoning_effort low, high, or max on K3 and K3.1 Pro, where the API default is max and thinking cannot be turned off. with_thinking(true) and with_thinking(false) send the thinking toggle on K2.6. K2.7 Code always thinks and offers no setting. The thinking keep field for K2.6 needs provider options.","sources":["s1049","s1050","s1048"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"resolve_effort"},{"version":"v2","path":"lib/ruby_llm/thinking.rb","method":"preferred_effort"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/moonshot_spec.rb","example":"RubyLLM::Providers::Moonshot chat turns thinking on and off on models with a toggle","notes":"Asserts thinking type disabled on kimi-k2.6."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/moonshot_spec.rb","example":"RubyLLM::Providers::Moonshot chat sets the effort of models that always think","notes":"Asserts reasoning_effort max on kimi-k3."}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"Responses web search on kimi-k2.6 returned annotations: [] on output_text, with the sources under web_search_call.action.sources, which is not inline citation. The Chat Completions reference documents none. RubyLLM parses annotations only from Chat Completions responses.","sources":["s1039","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_annotations"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"kimi-k2.6","notes":"POST /responses with web_search returned output_text annotations: [] and the sources under action.sources."}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"with_tools sends function tools and runs the tool loop, including in streaming conversations.","sources":["s1052","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling moonshot/kimi-k3 can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_moonshot_kimi-k3_can_use_tools.yml","notes":"Tool call and result turn."},{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling moonshot/kimi-k3 can use tools with multi-turn streaming conversations","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_moonshot_kimi-k3_can_use_tools_with_multi-turn_streaming_conversations.yml","notes":"Streamed tool calls across turns."}]},"tool_choice":{"v1":"missing","v2":"native","offered":true,"notes":"with_tool_options(choice:) sends auto, none, required, or a named function. Forcing a named function returns a 400 while thinking is on. Kimi K3 takes it with thinking turned off through with_provider_options, because the registry lists no off control for with_thinking(false), and takes required with thinking on. The reference says K2.6 and K2.7 Code reject required.","sources":["s1053","s1049"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/providers/moonshot/capabilities.rb","method":"augment"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control moonshot/kimi-k3 respects choice: :required for unrelated queries","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_moonshot_kimi-k3_respects_choice_required_for_unrelated_queries.yml","notes":"required with thinking on returned a weather tool call."},{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat tool choice and calls control moonshot/kimi-k3 respects specific tool choice","cassette":"spec/fixtures/vcr_cassettes/chat_tool_choice_and_calls_control_moonshot_kimi-k3_respects_specific_tool_choice.yml","notes":"Named function with thinking disabled."}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"The model can return several tool_calls in one response, and RubyLLM runs them all.","sources":["s1052","s1054"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling moonshot/kimi-k3 can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_moonshot_kimi-k3_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"Three dice_roll calls in one response."},{"status":"passed","recorded_on":"2026-11-06","model":"kimi-k3","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling moonshot/kimi-k3 can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_moonshot_kimi-k3_can_use_parallel_tool_calls.yml","notes":"Two different tools in one response."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"parallel_tool_calls is not honored: with false, kimi-k2.6 and kimi-k3 still returned four calls in one response on Chat Completions, and kimi-k3 on Responses, the same as the default. The field is accepted without validation and the request schemas do not document it. Nothing limits the calls, so with_tool_options(calls: :one) raises ArgumentError instead of sending a field that does nothing, and calls: :many leaves it off.","sources":["s1039","s1051","s1052"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_payload"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"kimi-k3","notes":"Four requests with parallel_tool_calls: false returned four tool calls. kimi-k2.6 and a Responses request did the same."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/moonshot_spec.rb","example":"RubyLLM::Providers::Moonshot calls: :one raises before sending, because the API ignores parallel_tool_calls","notes":"calls: :one raises ArgumentError and no request goes out."}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends the Responses web_search tool on K3.1 Pro and K3, with protocol: :responses, because Chat Completions has no server-run search. Sources come from action.sources; Kimi returns no inline citations. The moonshot/web-search official tool is an application-run function and has no RubyLLM alias.","sources":["s1051","s1054"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/responses.rb","method":"SERVER_TOOL_ALIASES, render_item, parse_server_tool_use"},{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with moonshot/kimi-k3 through Responses searches and lists the sources it used","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_moonshot_kimi-k3_through_responses_searches_and_lists_the_sources_it_used.yml","notes":"web_search_call items, sources under action.sources, and a counted search."},{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with moonshot/kimi-k3 through Responses streams a search and replays it so the conversation can continue","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_moonshot_kimi-k3_through_responses_streams_a_search_and_replays_it_so_the_conversation_can_continue.yml","notes":"Streamed search, then a second turn built from the local transcript."}]},"server_web_fetch":{"v1":"missing","v2":"missing","offered":true,"notes":"The moonshot/fetch official tool extracts a URL as Markdown, but the Responses API runs only web_search on the server. Fetch is a function declaration that the application runs through the Formula fibers endpoint, so it does not fit with_provider_tools. An integration needs a new client-run tool, which is a public API decision.","sources":["s1054"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"server_tool_aliases"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool types are function, custom (apply_patch only), namespace, and web_search. The official tool list has an Excel and CSV analysis tool but no general code execution tool.","sources":["s1046","s1047","s1051","s1054"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool types and the official tool list contain no file search tool. File questions go through text extraction into messages.","sources":["s1046","s1047","s1051","s1054"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API says other tool types are not supported, so it has no remote MCP tool. MCP servers are documented only for Kimi Playground, not as an API tool.","sources":["s1046","s1047","s1051","s1055"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"Kimi caches matching request prefixes automatically, with a 5m default. RubyLLM reads prompt_tokens_details cached_tokens and cache_write_tokens into cache_read and cache_write. Cache Write applies to K3 and K3.1 Pro, not K2.7 Code or K2.6.","sources":["s1056"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens, cache_write_tokens"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/moonshot_spec.rb","example":"RubyLLM::Providers::Moonshot chat parses reasoning and counts cache reads and writes","notes":"Asserts cached and written tokens leave the input bucket."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"Caching matches prefixes with no explicit markers. prompt_cache_options.mode supports only implicit, and the Messages API ignores cache_control markers inside messages. RubyLLM sends no breakpoints for this provider.","sources":["s1046","s1047","s1056","s1057"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"openai_prompt_caching?"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no cache resource endpoints. The caching guide says entries cannot be cleared manually and expire after inactivity.","sources":["s1046","s1047","s1056"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"The audited API documents no compaction or context management control. The multi-turn guide leaves context length to the application, because the API is stateless.","sources":["s1046","s1047","s1058"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"Chat#count_tokens posts the rendered messages, tools and response_format to /v1/tokenizers/estimate-token-count and returns data.total_tokens. The endpoint is an estimate. It counts tools and schemas though its reference lists only messages.","sources":["s1059"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"count_tokens_url, render_count_tokens_payload, parse_count_tokens_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_count_tokens_spec.rb","example":"RubyLLM::Chat#count_tokens with moonshot/kimi-k3 counts instructions, tools and structured output","cassette":"spec/fixtures/vcr_cassettes/chat_count_tokens_with_moonshot_kimi-k3_counts_instructions_tools_and_structured_output.yml","notes":"Tools and the schema raise the estimate from POST /v1/tokenizers/estimate-token-count."}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no image generation operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no image editing operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no video generation operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no speech generation operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no transcription operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no streaming transcription operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no speaker identification operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no moderation operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"ocr":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.ocr uploads the file with purpose file-extract, reads back the Markdown as one page, and deletes the upload. Kimi's extraction takes no model, so model: only labels the result, and any Kimi model id selects the provider. pages: raises ArgumentError because the endpoint extracts whole documents. Images are not supported for extraction.","sources":["s1045","s1060","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/documents.rb","method":"ocr"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/ocr_spec.rb","example":"RubyLLM::OCR basic functionality moonshot/kimi-k3 extracts markdown from a PDF","cassette":"spec/fixtures/vcr_cassettes/ocr_basic_functionality_moonshot_kimi-k3_extracts_markdown_from_a_pdf.yml","notes":"PDF uploaded with purpose file-extract, Markdown read back, upload deleted."}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no text embeddings operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no multimodal embeddings operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no reranking operation. Its endpoints cover chat, Responses, Messages, models, token estimation, tools, files, and batches.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to /v1/files, up to 100 MiB per file. purpose defaults to image, video or file-extract by content type, and batch JSONL uploads pass batch. expires_in raises ArgumentError because Kimi has no retention setting. Uploaded images are reusable in chat as ms:// references; uploaded documents are read as extracted text.","sources":["s1045"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/moonshot/files.rb","method":"render_upload_fields, parse_file_response"},{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads moonshot uploads a PDF through the Files API","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_moonshot_uploads_a_pdf_through_the_files_api.yml","notes":"POST /v1/files with the default file-extract purpose; byte size matches."},{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads moonshot/kimi-k3 reuses an uploaded file in chat and downloads its text","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_moonshot_kimi-k3_reuses_an_uploaded_file_in_chat_and_downloads_its_text.yml","notes":"Uploaded PDF used in a chat, then its extracted text downloaded."},{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads moonshot/kimi-k3 reads an uploaded image by reference","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_moonshot_kimi-k3_reads_an_uploaded_image_by_reference.yml","notes":"Uploaded image sent as an ms:// reference."}]},"file_download":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.download reads /v1/files/{id}/content. Kimi wraps extracted text in a JSON object with a content field, which RubyLLM unwraps to the text; batch results come back as written.","sources":["s1060","s1061"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/moonshot/files.rb","method":"download, extracted_text"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/uploaded_file_spec.rb","example":"RubyLLM::UploadedFile live uploads moonshot/kimi-k3 reuses an uploaded file in chat and downloads its text","cassette":"spec/fixtures/vcr_cassettes/uploadedfile_live_uploads_moonshot_kimi-k3_reuses_an_uploaded_file_in_chat_and_downloads_its_text.yml","notes":"Uploaded PDF used in a chat, then its extracted text downloaded."}]},"chat_batches":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.batch uploads JSONL and creates batches for /v1/chat/completions on K3.1 Pro, K3, K2.7 Code and K2.6, reads results back by custom_id, and costs 60 percent of standard rates. One model per batch; Kimi fixes temperature and top_p on these models.","sources":["s1061","s1062","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol, batch_cost_multiplier"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/batches.rb"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"kimi-k2.6","spec":"spec/ruby_llm/batch_spec.rb","example":"RubyLLM::Batch with moonshot/kimi-k2.6 answers a staged chat and appends the answer to its conversation","cassette":"spec/fixtures/vcr_cassettes/batch_with_moonshot_kimi-k2_6_answers_a_staged_chat_and_appends_the_answer_to_its_conversation.yml","notes":"One request submitted, completed in under a minute, and collected back into the chat."}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"Create Batch accepts only /v1/chat/completions, and the API has no embeddings endpoint.","sources":["s1046","s1047","s1062"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no advisor tool operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no agent API operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no agent skills API operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no asynchronous research job operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The Responses object reports background as always false, and the inventory has no response retrieval or cancellation endpoint.","sources":["s1046","s1047","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API lists function, custom (apply_patch only), namespace, and web_search as its tool types. No browser tool is documented.","sources":["s1046","s1047","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no classification operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API lists function, custom (apply_patch only), namespace, and web_search as its tool types. No computer tool is documented.","sources":["s1046","s1047","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The audited API documents no context editing control. The application owns its message history.","sources":["s1046","s1047","s1058"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no contextual embeddings operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no transcription operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no dubbing operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"fast_inference":{"v1":"missing","v2":"native","offered":true,"notes":"High-speed variants are ordinary model ids: kimi-k3.1-pro-highspeed and kimi-k2.7-code-highspeed. Both are in the registry and selected with model:. The service_tier field appears only on the Responses response object and is not a documented request option.","sources":["s1040","s1063","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/models.json"}]},"file_management":{"v1":"missing","v2":"missing","offered":true,"notes":"GET /v1/files lists files, and DELETE /v1/files/{file_id} deletes one. RubyLLM has no public verb for listing or deleting files on any provider.","sources":["s1064","s1065"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no file search store operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no completions endpoint, and the Chat Completions reference documents no suffix parameter.","sources":["s1046","s1047","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no fine-tuning operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no Google Maps grounding operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no guardrail configuration operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The API is stateless. The Responses object reports store and background as always false and previous_response_id and conversation as always null.","sources":["s1046","s1047","s1051","s1058"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no image batch operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no Interactions API operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"json_mode":{"v1":"missing","v2":"passthrough","offered":true,"notes":"response_format type json_object is documented. with_provider_options(response_format: { type: \"json_object\" }) forwards it; RubyLLM has no dedicated JSON object switch.","sources":["s1066","s1042"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no Bedrock Mantle operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The inventory has no compaction endpoint.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no music generation operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"Create Batch accepts only /v1/chat/completions.","sources":["s1046","s1047","s1062"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no omni video generation operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"Kimi serves its own models. The OpenAI and Anthropic compatible endpoints are formats for those models, not partner model routes.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"prefix_completion":{"v1":"missing","v2":"passthrough","offered":true,"notes":"Partial Mode continues from a final assistant message marked partial: true. RubyLLM does not render that field, but with_provider_options(messages: [...]) replaces the rendered messages and forwards it. There is no RubyLLM method for this.","sources":["s1067","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no programmatic tool calling operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_caching(key:, ttl:, mode:) sends prompt_cache_key and prompt_cache_options. ttl takes 5m or 1h, and mode supports only implicit.","sources":["s1056","s1039"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"prompt_cache_params"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/moonshot_spec.rb","example":"RubyLLM::Providers::Moonshot chat sends a prompt cache key and lifetime without breakpoints","notes":"Asserts the key and 1h lifetime, with no breakpoint."}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no provider-managed fallback operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no realtime session operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no response caching operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"responses":{"v1":"missing","v2":"native","offered":true,"notes":"protocol: :responses selects POST /v1/responses on K3.1 Pro and K3. It is stateless, with store always false, and tool_choice accepts only auto. Text travels as input_text parts because Kimi drops web search for string content. The custom apply_patch tool and the Anthropic-compatible /anthropic/v1/messages endpoint are not integrated.","sources":["s1051","s1057","s1046"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot/responses.rb"},{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with moonshot/kimi-k3 through Responses searches and lists the sources it used","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_moonshot_kimi-k3_through_responses_searches_and_lists_the_sources_it_used.yml","notes":"web_search_call items, sources under action.sources, and a counted search."},{"status":"passed","recorded_on":"2026-11-09","model":"kimi-k3","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with moonshot/kimi-k3 through Responses streams a search and replays it so the conversation can continue","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_moonshot_kimi-k3_through_responses_streams_a_search_and_replays_it_so_the_conversation_can_continue.yml","notes":"Streamed search, then a second turn built from the local transcript."}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"Create Batch accepts only /v1/chat/completions, so Responses requests cannot be batched.","sources":["s1046","s1047","s1062"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"search_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/tools/search and /v1/tools/search_pro return structured search results outside a chat request. RubyLLM has no standalone search operation.","sources":["s1068","s1069","s1070"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_apply_patch":{"v1":"missing","v2":"missing","offered":true,"notes":"The Responses API accepts one custom tool, apply_patch, with a Lark grammar, and returns custom_tool_call items. This provider declares only Chat Completions, so RubyLLM cannot reach it.","sources":["s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The Responses tool types include no image generation tool. web_search can return image search results, which is not generation.","sources":["s1046","s1047","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"Kimi has no hosted tool search. Its pattern is a search_tools function the application implements, plus tools injected mid-conversation through a system message or a Responses additional_tools item.","sources":["s1046","s1047","s1071","s1072","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no X search tool operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The Responses API says tool types other than function, custom (apply_patch only), namespace, and web_search are not supported.","sources":["s1046","s1047","s1051"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no sound effects operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no streaming speech operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no task budget operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no text analysis operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no vector store operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no video batch operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no video editing operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no video extension operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no voice cloning operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no voice conversion operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"web_fetch_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /v1/tools/fetch returns the title and Markdown body of a URL. RubyLLM has no standalone fetch operation.","sources":["s1073","s1070"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The audited Kimi API inventory documents no WebSocket operation.","sources":["s1046","s1047"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/moonshot.rb","method":"protocol"}]}},"gaps":[{"notes":"Web fetch tool: the moonshot/fetch official tool runs client-side through the Formula fibers endpoint, so it does not fit with_provider_tools.","sources":["s1054"]},{"notes":"Standalone search API, Standalone web fetch API: the /v1/tools endpoints have no RubyLLM operation.","sources":["s1068","s1073"]},{"notes":"File listing and deletion: Kimi documents both, but RubyLLM has no public verb for them.","sources":["s1064","s1065"]},{"notes":"Apply patch tool: the Responses custom apply_patch tool has no RubyLLM alias for this provider.","sources":["s1051"]}]},{"id":"typesafe","name":"TypeSafe","notes":"Covers TypeSafe's System One API at api.typesafe.ai, which has two endpoints: POST /v1/systemone and GET /v1/models. Its main value, typed judgments with probabilities, choices, and scores, has no row here: RubyLLM reaches it through RubyLLM.judge and RubyLLM::Judge, so most rows read as not in this API.","features":{"chat":{"v1":"na","v2":"na","offered":false,"notes":"Jev does not generate text or hold a conversation, and the API has no chat endpoint. RubyLLM registers TypeSafe models as judgment models, so a chat cannot reach this provider.","sources":["s1074","s1075","s1076","s1077","s1078","s1079"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"streaming":{"v1":"na","v2":"na","offered":false,"notes":"Jev does not stream text. POST /v1/systemone returns one JSON response.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"structured_output":{"v1":"na","v2":"na","offered":false,"notes":"The API has no schema-based generation. Its answers are always typed by question type, which RubyLLM exposes through judgments rather than chat.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"vision":{"v1":"na","v2":"na","offered":false,"notes":"The API has no chat, so it has no chat attachments. Jev does read PNG and JPEG images inside System One requests. That judgment path is a preview, and RubyLLM rejects attachments sent to it.","sources":["s1074","s1075","s1076","s1077","s1080","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"pdf_input":{"v1":"na","v2":"na","offered":false,"notes":"The API has no chat attachments, and the Models page lists only text and images as Jev input.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"The API has no chat attachments. The Models page states Jev does not natively accept audio.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"video_input":{"v1":"na","v2":"na","offered":false,"notes":"The API has no chat attachments. The Models page states Jev does not natively accept video.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"thinking":{"v1":"na","v2":"na","offered":false,"notes":"The System One response carries answers, the model, and input and output token usage. It documents no thinking content.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"thinking_controls":{"v1":"na","v2":"na","offered":false,"notes":"The System One request takes model, state, questions, and assets. It documents no thinking settings.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"citations":{"v1":"na","v2":"na","offered":false,"notes":"The System One response documents no citations. The citation cookbook checks citations supplied by an application with a Choice question.","sources":["s1074","s1075","s1076","s1077","s1082"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"tools":{"v1":"na","v2":"na","offered":false,"notes":"The request has no tool definitions and the response has no tool calls. The function calling cookbook maps function names and arguments onto ordinary questions in application code.","sources":["s1074","s1075","s1076","s1077","s1083","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"tool_choice":{"v1":"na","v2":"na","offered":false,"notes":"The API has no tool definitions, so it has no tool choice.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"parallel_tools":{"v1":"na","v2":"na","offered":false,"notes":"The API returns no tool calls. Several questions in one request are not tool calls.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"The API returns no tool calls, so it has no limit on them.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_web_search":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including web search.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including web fetch.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including code execution.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_file_search":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including file search.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including remote MCP.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"prompt_caching":{"v1":"na","v2":"na","offered":false,"notes":"The API reference and Models page document no prompt caching. Caches in the cookbooks are local helpers in the example code.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no cache markers.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no cache resources.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"The API is stateless per request and documents no context controls that reduce it. Long context raises the limits instead.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"token_counting":{"v1":"na","v2":"na","offered":false,"notes":"The API has no token counting endpoint. Each response reports input and output tokens after the judgment runs.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers image generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers image editing.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers video generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers speech generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"transcription":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers transcription.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"streaming_transcription":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers streaming transcription.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers speaker identification.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The API has no moderation endpoint. The guardrails cookbook builds screening from Noul and Score questions, which RubyLLM reaches through judgments.","sources":["s1074","s1075","s1076","s1077","s1084"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"ocr":{"v1":"na","v2":"na","offered":false,"notes":"The API has no OCR or document parsing endpoint. The cookbooks read scanned pages by asking questions about images, and Box questions locate regions but do not return text.","sources":["s1074","s1075","s1076","s1077","s1085","s1086"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither returns embeddings.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither returns embeddings.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The API has no rerank endpoint. The re-ranking cookbook asks one question per query and candidate pair in application code.","sources":["s1074","s1075","s1076","s1077","s1087"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"file_upload":{"v1":"na","v2":"na","offered":false,"notes":"The API has no file store. Images travel inside a single System One request, as base64 or multipart parts.","sources":["s1074","s1075","s1076","s1077","s1080"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"The API has no file store to download from.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API has no chat and no batch endpoint. Several questions in one request are not a batch job.","sources":["s1074","s1075","s1076","s1077","s1088"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API has no embeddings and no batch endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including an advisor.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row covers a Perplexity product. TypeSafe has no equivalent endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The API has no skills parameter. The Agent skill page offers instructions for coding agents that write TypeSafe code, which is not an API operation.","sources":["s1074","s1075","s1076","s1077","s1089"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no asynchronous jobs. The asynchronous Python client is a client-side interface to the same requests.","sources":["s1074","s1075","s1076","s1077","s1090"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no background jobs.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no browser tool.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"classification":{"v1":"missing","v2":"native","offered":true,"notes":"Choice questions select one label from a defined set and return a probability for every label, with confidence. RubyLLM exposes them as `choice` on RubyLLM::Judge and RubyLLM.judge, through POST /v1/systemone. The API has no separate classify endpoint.","sources":["s1074","s1091","s1092","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/system_one/judgments.rb","method":"render_judgment_payload, render_question"},{"version":"v2","path":"lib/ruby_llm/protocols/system_one/responses.rb","method":"parse_choice, parse_probabilities"},{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"system_one"}],"verification":[{"status":"passed","recorded_on":"2026-10-11","model":"jev-latest","spec":"spec/ruby_llm/providers/typesafe_spec.rb","example":"RubyLLM::Providers::TypeSafe with the TypeSafe API judges all three question types through the compact DSL","cassette":"spec/fixtures/vcr_cassettes/providers_typesafe_with_the_typesafe_api_judges_all_three_question_types_through_the_compact_dsl.yml","notes":"A Choice question over three departments returns the selected label, a probability per label, and confidence. The alias jev-latest answered as jev-1.13.0."}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no computer use tool.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"context_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API is stateless per request and documents no context editing.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither returns embeddings.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"dedicated_transcription_api":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers transcription.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers dubbing.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"fast_inference":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no fast tier or speed setting. Long context is the only documented request mode, and it is slower.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"The API has no file store to list or delete from.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The API has no file search stores.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no text completion endpoint.","sources":["s1074","s1075","s1076","s1077","s1079"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The Models page states Jev is not fine-tuned or adapted with customer data, and the same weights serve every account.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no grounding tools.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The API has no guardrail configuration. The guardrails cookbook applies thresholds in application code.","sources":["s1074","s1075","s1076","s1077","s1084"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"hosted_conversations":{"v1":"na","v2":"na","offered":false,"notes":"The API keeps no conversation. Each request carries its own state.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API has no image generation and no batch endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"This row covers a Google protocol. TypeSafe does not document it.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"json_mode":{"v1":"na","v2":"na","offered":false,"notes":"The API has no free-form generation to constrain. Answers are always typed JSON.","sources":["s1074","s1075","s1076","s1077","s1079"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row covers an AWS Bedrock route. TypeSafe does not document it.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no compaction operation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers music generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API has no batch endpoint. Several questions in one request are not a batch job.","sources":["s1074","s1075","s1076","s1077","s1088"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers video generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"TypeSafe serves its own Jev models through one endpoint and documents no partner model protocols.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no text completion endpoint.","sources":["s1074","s1075","s1076","s1077","s1079"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no tool calling.","sources":["s1074","s1075","s1076","s1077","s1078"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no prompt cache options.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no fallback controls. An alias such as jev-latest points to one versioned model.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no realtime sessions.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no response caching. Caches in the cookbooks are local helpers in the example code.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"responses":{"v1":"na","v2":"na","offered":false,"notes":"The API does not document the Responses protocol.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API documents neither the Responses protocol nor a batch endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"search_api":{"v1":"na","v2":"na","offered":false,"notes":"The API has no search endpoint. The line-by-line search cookbook builds search from questions in application code.","sources":["s1074","s1075","s1076","s1077","s1093"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including apply patch.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including image generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no tools, so it has no tool search.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no provider-run tools, including X search.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no shell or computer tools.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers sound effects.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers speech generation.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no task budgets. Limits apply per request and per question through context length.","sources":["s1074","s1075","s1076","s1077","s1081"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The API has no dedicated text analysis endpoint for intent, sentiment, or topics. An application asks its own questions through judgments.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The API has no vector stores.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The API has no video generation and no batch endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers video editing.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers video extension.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers voice cloning.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The API documents two endpoints, POST /v1/systemone and GET /v1/models. Neither offers voice conversion.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"web_fetch_api":{"v1":"na","v2":"na","offered":false,"notes":"The API has no web fetch endpoint.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no WebSocket transport.","sources":["s1074","s1075","s1076","s1077"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/typesafe.rb","method":"protocol, system_one"}]}},"gaps":[{"notes":"Judgment images, Box questions: the API accepts PNG and JPEG images in a request and answers Box questions on Jev 1.14, both as a preview behind the TypeSafe-Danger-Zone-Images header. RubyLLM raises UnsupportedAttachmentError for attachments, has no Box question type, and cannot send the header from a judgment.","sources":["s1080","s1086","s1094"]},{"notes":"Long context: the TypeSafe-Danger-Zone-Long-Context header raises both context limits to 262k tokens. A judgment has no way to send a request header, so RubyLLM stays within the default limits.","sources":["s1081"]},{"notes":"Judgment cost: TypeSafe publishes a per-token input price, but the registry holds no price for the jev-latest and jev-preview aliases, so a hosted judgment reports token counts and a nil cost. Versioned model IDs are absent from the listing and need assume_model_exists.","sources":["s1081"]}]},{"id":"zai","name":"Z.ai","notes":"This covers the Z.ai API documented at docs.z.ai: chat, media, tool and agent endpoints under api.z.ai. RubyLLM speaks a Chat Completions dialect at /api/paas/v4, so media operations reach Z.ai only where that shared protocol matches. Setting zai_api_base can point it at BigModel or the GLM Coding Plan endpoint; BigModel documents its platform separately, outside this comparison.","features":{"chat":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM sends messages to POST /paas/v4/chat/completions through its Z.ai dialect of Chat Completions. Instructions keep the system role, and with_end_user sends user_id.","sources":["s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol, api_base"},{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"render_role, apply_end_user"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality zai/glm-5.3-flash can have a basic conversation","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_zai_glm-5_3-flash_can_have_a_basic_conversation.yml","notes":"POST /paas/v4/chat/completions returns content and usage."},{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_spec.rb","example":"RubyLLM::Chat basic chat functionality zai/glm-5.3-flash can handle multi-turn conversations","cassette":"spec/fixtures/vcr_cassettes/chat_basic_chat_functionality_zai_glm-5_3-flash_can_handle_multi-turn_conversations.yml","notes":"Earlier turns replay from the local transcript."}]},"streaming":{"v1":"missing","v2":"native","offered":true,"notes":"stream: true returns server-sent chunks, parsed into RubyLLM chunks with usage on the final event. RubyLLM does not send tool_stream; a recording shows streamed tool calls without it, and with_provider_options can set it.","sources":["s1096","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_streaming_spec.rb","example":"RubyLLM::Chat streaming responses zai/glm-5.3-flash supports streaming responses","cassette":"spec/fixtures/vcr_cassettes/chat_streaming_responses_zai_glm-5_3-flash_supports_streaming_responses.yml","notes":"Text deltas and usage arrive over SSE."},{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling zai/glm-5.3-flash can use tools with multi-turn streaming conversations","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_zai_glm-5_3-flash_can_use_tools_with_multi-turn_streaming_conversations.yml","notes":"Streamed tool calls without tool_stream."}]},"structured_output":{"v1":"missing","v2":"native","offered":true,"notes":"with_schema sends response_format with type json_schema. The docs describe only json_object, for text models, and do not describe json_schema or say whether the schema is enforced. A recording on GLM-5.3-Flash returns JSON matching the schema.","sources":["s1097","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"render_payload, schema_strict"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_schema_spec.rb","example":"RubyLLM::Chat #with_schema with zai/glm-5.3-flash accepts a JSON schema and returns structured output","cassette":"spec/fixtures/vcr_cassettes/chat_with_schema_with_zai_glm-5_3-flash_accepts_a_json_schema_and_returns_structured_output.yml","notes":"A strict json_schema request returns matching JSON."}]},"vision":{"v1":"missing","v2":"native","offered":true,"notes":"Image attachments render as image_url parts, by URL or Base64 data URL. Vision models include the GLM-5.3-Flash series and the GLM-4.6V family.","sources":["s1095","s1098"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"render_content, supported_attachment?"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/media.rb","method":"render_image"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat vision models zai/glm-5.3-flash can understand local images","cassette":"spec/fixtures/vcr_cassettes/chat_vision_models_zai_glm-5_3-flash_can_understand_local_images.yml","notes":"A local image goes out as a data URL and the model describes it."}]},"pdf_input":{"v1":"missing","v2":"native","offered":true,"notes":"ask(with:) reads a PDF on the GLM-5.3-Flash series, the only models that accept one. An inline PDF uploads through the Files API and attaches by file_id, a file uploaded with RubyLLM.upload attaches the same way, and a PDF URL goes through as file_url. auto_upload_large_files: false turns the upload off and raises UnsupportedAttachmentError, as does any other model. Text files are inlined as text parts.","sources":["s1095","s1098"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"supported_attachment?, render_content, remote_pdf?"},{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"provider_file_references?, default_large_file_upload_threshold, provider_file_attachable?"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"preprocess_attachment"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/protocols/zai/files_spec.rb","example":"RubyLLM::Protocols::ZAI::Files uploads an inline PDF on its own for a model that reads file ids","cassette":"spec/fixtures/vcr_cassettes/protocols_zai_files_uploads_an_inline_pdf_on_its_own_for_a_model_that_reads_file_ids.yml","notes":"An inline PDF was uploaded and answered about by file_id."},{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/protocols/zai/files_spec.rb","example":"RubyLLM::Protocols::ZAI::Files reads a PDF from a public URL","cassette":"spec/fixtures/vcr_cassettes/protocols_zai_files_reads_a_pdf_from_a_public_url.yml","notes":"Z.ai fetched and read the PDF URL."}]},"audio_input":{"v1":"na","v2":"na","offered":false,"notes":"No audio request shape works: input_audio returned 400 (code 1210) on glm-5.3-flash, glm-5v-turbo and glm-4.5v, audio_url returned 400 (code 1214) on glm-5.3-flash, glm-5v-turbo and glm-4.6v, and glm-4.6v accepted input_audio and returned empty content. The content schema defines only text, image_url, video_url and file parts. RubyLLM raises UnsupportedAttachmentError for audio attachments on this provider.","sources":["s1095","s1098"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"supported_attachment?, render_content"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI attachments rejects documents and audio","notes":"Audio attachments raise before a request."},{"status":"unavailable","checked_on":"2026-11-09","model":"glm-5.3-flash","notes":"Spoken-word WAV through input_audio returned 400 invalid parameter on glm-5.3-flash and glm-4.5v, and empty content on glm-4.6v. A header-only WAV returned the same 400 on glm-5v-turbo."}]},"video_input":{"v1":"missing","v2":"native","offered":true,"notes":"The GLM-5.3-Flash series takes video attachments as video_url parts, by URL for remote videos and as a data URI for local ones. Z.ai documents MP4, MKV and MOV up to 200 MB and describes only URLs, but a data URI was accepted live. Uploaded videos are not rendered as video_url.","sources":["s1095","s1098"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"supported_attachment?, render_content, render_video"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models zai/glm-5.3-flash can understand local videos","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_zai_glm-5_3-flash_can_understand_local_videos.yml","notes":"Data URI in video_url."},{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_content_spec.rb","example":"RubyLLM::Chat video models zai/glm-5.3-flash can understand remote videos without extension","cassette":"spec/fixtures/vcr_cassettes/chat_video_models_zai_glm-5_3-flash_can_understand_remote_videos_without_extension.yml","notes":"Remote URL in video_url."}]},"thinking":{"v1":"missing","v2":"native","offered":true,"notes":"reasoning_content is parsed into message.thinking, including on streams, and earlier reasoning replays as reasoning_content. Supported from the GLM-4.6 series up.","sources":["s1099","s1100","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"render_thinking"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"parse_thinking_text"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking zai/glm-5.3-flash returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_zai_glm-5_3-flash_returns_thinking_when_available.yml","notes":"reasoning_content appears as thinking text."},{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking zai/glm-5.3-flash streams thinking content when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_zai_glm-5_3-flash_streams_thinking_content_when_available.yml","notes":"Thinking deltas arrive on the stream."}]},"thinking_controls":{"v1":"missing","v2":"native","offered":true,"notes":"with_thinking sends thinking.type and reasoning_effort. Effort applies to GLM-5.2 and later; GLM-5.3 and GLM-5.3-Flash take only low, high and max, and return an error if thinking is disabled. thinking.clear_thinking has no shared name and goes through with_provider_options.","sources":["s1099","s1100","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"render_payload"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_thinking_spec.rb","example":"RubyLLM::Chat with extended thinking zai/glm-5.3-flash returns thinking when available","cassette":"spec/fixtures/vcr_cassettes/chat_with_extended_thinking_zai_glm-5_3-flash_returns_thinking_when_available.yml","notes":"The request carries thinking.type enabled and reasoning_effort low."},{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI chat turns thinking off on models with a toggle","notes":"with_thinking(false) sends thinking.type disabled and no effort."}]},"citations":{"v1":"missing","v2":"native","offered":true,"notes":"Sources of the web_search tool become Message#citations, with address, title, snippet and source index, in full and streamed responses. Z.ai returns them only with that tool, and it places no markers tying a source to a span of the answer, so citations have no offsets. with_citations alone asks for nothing.","sources":["s1101","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"parse_root_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"extract_citations"},{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/streaming.rb","method":"parse_chunk_citations"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with zai/glm-5.3-flash searches, cites the sources, and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_zai_glm-5_3-flash_searches_cites_the_sources_and_continues_from_the_local_transcript.yml","notes":"Citation values carry the link and title of each web_search result."},{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with zai/glm-5.3-flash streams a search and collects the citations from the last chunk","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_zai_glm-5_3-flash_streams_a_search_and_collects_the_citations_from_the_last_chunk.yml","notes":"Citations collected from the final streamed chunk."}]},"tools":{"v1":"missing","v2":"native","offered":true,"notes":"Function tools render as type function, with up to 128 functions per request. Tool results and multi-turn tool loops run through the shared Chat Completions tools module.","sources":["s1102","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"tool_for, parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling zai/glm-5.3-flash can use tools","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_zai_glm-5_3-flash_can_use_tools.yml","notes":"A tool call runs and its result returns to the model."},{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling zai/glm-5.3-flash can use tools with multi-turn streaming conversations","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_zai_glm-5_3-flash_can_use_tools_with_multi-turn_streaming_conversations.yml","notes":"Tool loop over a streamed response."}]},"tool_choice":{"v1":"missing","v2":"native","offered":true,"notes":"with_tool_options(choice: :auto) sends tool_choice. Z.ai documents auto as the only accepted value, so other choices are left for the API to answer.","sources":["s1102","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"build_tool_choice"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_tool_options"}]},"parallel_tools":{"v1":"missing","v2":"native","offered":true,"notes":"Recordings show several tool calls in one response, with different tools and with the same tool. The docs do not describe this behaviour.","sources":["s1102","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/tools.rb","method":"parse_tool_calls"}],"verification":[{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling zai/glm-5.3-flash can use parallel tool calls","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_zai_glm-5_3-flash_can_use_parallel_tool_calls.yml","notes":"One response carries weather and best_language_to_learn calls."},{"status":"passed","recorded_on":"2026-10-20","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_tools_spec.rb","example":"RubyLLM::Chat function calling zai/glm-5.3-flash can handle multiple tool calls in a single response","cassette":"spec/fixtures/vcr_cassettes/chat_function_calling_zai_glm-5_3-flash_can_handle_multiple_tool_calls_in_a_single_response.yml","notes":"One response carries three dice_roll calls."}]},"parallel_tool_control":{"v1":"na","v2":"na","offered":false,"notes":"parallel_tool_calls is not honored: with false, glm-5.3-flash returned four calls in one response, the same as the default, and unknown fields are also accepted. The Chat Completion schema and Function Calling guide do not mention it. Nothing limits the calls, so with_tool_options(calls: :one) raises ArgumentError instead of sending a field that does nothing, and calls: :many leaves it off.","sources":["s1102","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"render_payload"},{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_tool_options"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"glm-5.3-flash","notes":"Four requests with parallel_tool_calls: false returned four tool calls, the same as the default and true."},{"status":"unit","checked_on":"2026-11-09","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI calls: :one raises before sending, because the API ignores parallel_tool_calls","notes":"calls: :one raises ArgumentError and no request goes out."}]},"server_web_search":{"v1":"missing","v2":"native","offered":true,"notes":"with_provider_tools(:web_search) sends the Chat tool of type web_search with search_result on; its options, such as count or search_engine, go inside the web_search key. It sits next to function tools. Results come back as Message#citations. Z.ai reports no search count in usage, so tokens.server_tool_use stays empty.","sources":["s1101","s1095","s1103","s1104"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"SERVER_TOOL_ALIASES, parse_root_citations"},{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"server_tool_aliases"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"merge_server_tool_entries"}],"verification":[{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with zai/glm-5.3-flash searches, cites the sources, and continues from the local transcript","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_zai_glm-5_3-flash_searches_cites_the_sources_and_continues_from_the_local_transcript.yml","notes":"Top-level web_search array, an answer from it, and a second turn built from local history."},{"status":"passed","recorded_on":"2026-11-10","model":"glm-5.3-flash","spec":"spec/ruby_llm/chat_provider_tools_spec.rb","example":"RubyLLM::Chat web search with zai/glm-5.3-flash streams a search and collects the citations from the last chunk","cassette":"spec/fixtures/vcr_cassettes/chat_web_search_with_zai_glm-5_3-flash_streams_a_search_and_collects_the_citations_from_the_last_chunk.yml","notes":"The web_search array arrives on the final chunk."}]},"server_web_fetch":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool list has no fetch tool. Web Reader is a separate endpoint, recorded under the standalone web fetch row.","sources":["s1105","s1106","s1107"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_code_execution":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema lists function, retrieval and web_search tools. No code execution tool is documented.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_file_search":{"v1":"na","v2":"unknown","offered":null,"notes":"The Chat Completion schema defines a retrieval tool with knowledge_id and prompt_template, while the same page says only functions are supported as tools. No guide or knowledge base API is documented, so a real knowledge base cannot be created. A live request with a made-up knowledge_id was accepted and stopped at the token limit in reasoning, with no retrieval output, which does not settle whether a real one would work. If it works, with_provider_options(tools: [...]) forwards it.","sources":["s1095","s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"glm-5.3-flash","notes":"A retrieval tool with a made-up knowledge_id returned 200 with reasoning only, finish_reason length."}]},"server_mcp":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no remote MCP tool type. Z.ai hosts MCP servers for MCP clients to connect to themselves, which is a client integration and not a tool the model API runs.","sources":["s1105","s1106","s1095","s1108"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"prompt_caching":{"v1":"missing","v2":"native","offered":true,"notes":"Z.ai caches repeated context automatically, with no request option, and the FAQ describes the mechanism as an open beta. RubyLLM reads usage.prompt_tokens_details.cached_tokens as cache_read and removes those tokens from input.","sources":["s1109","s1110","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/chat.rb","method":"cache_read_tokens, input_tokens"},{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"openai_prompt_caching?"}],"verification":[{"status":"unit","checked_on":"2026-11-06","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI chat parses reasoning and counts cached input","notes":"cached_tokens map to cache_read."}]},"cache_boundaries":{"v1":"na","v2":"na","offered":false,"notes":"The Context Caching guide describes implicit caching with no manual configuration, and the chat schema has no cache markers.","sources":["s1105","s1106","s1109"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"explicit_cache":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no cache resources to create or manage. Caching is implicit, and the pricing page lists cached input storage as free for a limited time.","sources":["s1105","s1106","s1109","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"compaction":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no summarization or sliding window control. thinking.clear_thinking only removes earlier reasoning and is recorded under context editing.","sources":["s1105","s1106","s1095","s1100"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"token_counting":{"v1":"missing","v2":"native","offered":true,"notes":"chat.count_tokens posts the messages and tools to POST /paas/v4/tokenizer and reads usage.total_tokens, which includes image tokens. The tokenizer has no field for response format or thinking, so those aren't counted. It has no token ids, so RubyLLM.tokenize still raises an error. The schema lists glm-4.6, and glm-5.1 and glm-5.3-flash also answered.","sources":["s1111"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/token_counting.rb","method":"render_count_tokens_payload"},{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"ChatCompletions"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"glm-5.3-flash","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI token counting counts instructions, tools and images before generation","cassette":"spec/fixtures/vcr_cassettes/providers_zai_token_counting_counts_instructions_tools_and_images_before_generation.yml","notes":"Counts grow with instructions and a tool, and with an attached image, against the live tokenizer."}]},"image_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.paint posts to images/generations through the inherited Chat Completions image seams and reads the returned image URL. GLM-Image is not in the registry, so calls need provider: :zai and assume_model_exists: true. RubyLLM also sends n, which the schema does not list, and a null size when none is given; no recording covers this provider. The asynchronous image route has no integration.","sources":["s1112","s1113","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/images.rb","method":"render_image_payload, parse_image_responses"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"paint"}]},"image_editing":{"v1":"na","v2":"na","offered":false,"notes":"The image endpoints take a model, prompt, quality and size only, and no edit endpoint or input image is documented.","sources":["s1105","s1106","s1112"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"video_generation":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.animate with assume_model_exists: true posts to /paas/v4/videos/generations and polls /paas/v4/async-result. with: sends a first frame, or a first and last frame, as URLs or data URIs. quality, with_audio, size, fps and duration go through provider_options. The finished video stays at Z.ai's URL. cogvideox-3 is not in the model catalog.","sources":["s1114","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/videos.rb","method":"render_video_payload"},{"version":"v2","path":"lib/ruby_llm/providers/zai/videos.rb","method":"parse_video_job_status"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"cogvideox-3","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI video generation generates a video and points at the hosted file","cassette":"spec/fixtures/vcr_cassettes/providers_zai_video_generation_generates_a_video_and_points_at_the_hosted_file.yml","notes":"A five second speed-mode clip finished; the hosted MP4 answered with video/mp4."}]},"speech":{"v1":"na","v2":"na","offered":false,"notes":"The SDK pages list this as a feature, but the docs index, OpenAPI file and pricing page show no endpoint or model for it, and GET /models lists no model for it. The live route answered 400 Unknown Model for the ids tried. RubyLLM.speak would post to audio/speech through the inherited protocol.","sources":["s1104","s1105","s1106","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"speech_url"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"cogtts","notes":"POST /audio/speech with cogtts and glm-tts returned 400 code 1211: Unknown Model."}]},"transcription":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.transcribe posts the file and model to audio/transcriptions through the inherited protocol and reads text. GLM-ASR-2512 is not in the registry, so calls need provider: :zai and assume_model_exists: true. Z.ai limits audio to WAV or MP3, 30 seconds and 25 MB. hotwords goes through provider_options. No recording covers this provider.","sources":["s1115","s1116","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"render_transcription_payload, parse_transcription_response"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"transcribe"}]},"streaming_transcription":{"v1":"missing","v2":"native","offered":true,"notes":"A block on RubyLLM.transcribe sends stream=true and reads the documented transcript.text.delta and transcript.text.done events into TranscriptionChunk objects. This is server-sent events for a file, not a live session. No recording covers this provider.","sources":["s1115","s1116"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"stream_transcription, STREAM_TYPES"}]},"diarization":{"v1":"na","v2":"na","offered":false,"notes":"The transcription request and response have no speaker fields.","sources":["s1105","s1106","s1115"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"moderation":{"v1":"na","v2":"na","offered":false,"notes":"The SDK pages list this as a feature, but the docs index, OpenAPI file and pricing page show no endpoint or model for it, and GET /models lists no model for it. The live route answered 400 Unknown Model for the ids tried. The sensitive finish reason and error code 1301 flag blocked content and do not classify input on request. RubyLLM.moderate would post to moderations through the inherited protocol.","sources":["s1104","s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/moderation.rb","method":"moderation_url"},{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"FINISH_REASONS"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"moderation","notes":"POST /moderations returned 400 code 1211: Unknown Model."}]},"ocr":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.ocr with assume_model_exists: true posts URLs or data URIs to /paas/v4/layout_parsing and builds each page from layout_details, with images and tables. pages: takes one continuous range; other selections raise ArgumentError. glm-ocr is not in the model catalog.","sources":["s1117","s1118"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai/ocr.rb","method":"render_ocr_payload"},{"version":"v2","path":"lib/ruby_llm/providers/zai/ocr.rb","method":"parse_ocr_response"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"glm-ocr","spec":"spec/ruby_llm/providers/zai_spec.rb","example":"RubyLLM::Providers::ZAI OCR reads each page of a PDF","cassette":"spec/fixtures/vcr_cassettes/providers_zai_ocr_reads_each_page_of_a_pdf.yml","notes":"Reads the sample PDF as one page of Markdown with token usage."}]},"embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The SDK pages list this as a feature, but the docs index, OpenAPI file and pricing page show no endpoint or model for it, and GET /models lists no model for it. The live route answered 400 Unknown Model for the ids tried. RubyLLM.embed would post to embeddings through the inherited protocol.","sources":["s1104","s1105","s1106","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/embeddings.rb","method":"embedding_url"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"embedding-3","notes":"POST /embeddings with embedding-2 and embedding-3 returned 400 code 1211: Unknown Model."}]},"multimodal_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no multimodal embeddings.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"reranking":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no reranking endpoint. RubyLLM.rerank raises an error on this provider.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"file_upload":{"v1":"missing","v2":"native","offered":true,"notes":"RubyLLM.upload posts to /paas/v4/files with purpose, defaulting to user_data, and the resulting file attaches to chat by file_id on the GLM-5.3-Flash series. A live check found purpose agent files fail to parse in chat and user_data takes PDFs and Office files only. Z.ai has no route to read a file back, so UploadedFile.find and download raise an error. expires_in and uri are rejected.","sources":["s1119","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/zai/files.rb","method":"render_upload_fields"},{"version":"v2","path":"lib/ruby_llm/providers/zai/chat.rb","method":"document_attachments"},{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}],"verification":[{"status":"passed","recorded_on":"2026-11-09","model":"glm-5.3-flash","spec":"spec/ruby_llm/protocols/zai/files_spec.rb","example":"RubyLLM::Protocols::ZAI::Files uploads a PDF and answers from it in chat","cassette":"spec/fixtures/vcr_cassettes/protocols_zai_files_uploads_a_pdf_and_answers_from_it_in_chat.yml","notes":"Uploads the sample PDF, then answers a question about it through file_id."}]},"file_download":{"v1":"na","v2":"na","offered":false,"notes":"The file API documents upload only, with no route for reading file content back.","sources":["s1105","s1106","s1119"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"chat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint, and the pricing page lists no batch rate.","sources":["s1105","s1106","s1103"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"embedding_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"advisor_tool":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no advisor tool.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"agent_responses":{"v1":"na","v2":"na","offered":false,"notes":"This row is a Perplexity API. Z.ai agents are recorded under hosted conversations.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"agent_skills":{"v1":"na","v2":"na","offered":false,"notes":"The API documents no skills resource. Skills in the Coding Plan pages belong to the coding tools that run them.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"async_research":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no research job endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"background_responses":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no background job option for chat.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"browser_use":{"v1":"na","v2":"na","offered":false,"notes":"The GLM-5.3-Flash guide describes the model operating browsers inside agent environments. The chat schema has no browser tool.","sources":["s1105","s1106","s1098","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"classification":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no classification endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"computer_use":{"v1":"na","v2":"na","offered":false,"notes":"The GLM-5.3-Flash guide shows computer use prompts for agent environments. The chat schema has no computer tool.","sources":["s1105","s1106","s1098","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"context_editing":{"v1":"missing","v2":"passthrough","offered":true,"notes":"thinking.clear_thinking controls whether reasoning from earlier turns stays in the context, and it defaults to clearing. with_provider_options(thinking: {clear_thinking: false}) forwards it. It covers reasoning only. No tool result or message clearing is documented.","sources":["s1100","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"contextual_embeddings":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no embeddings endpoint of this kind.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"dedicated_transcription_api":{"v1":"missing","v2":"native","offered":true,"notes":"audio/transcriptions is a dedicated endpoint, and RubyLLM.transcribe reaches it. This overlaps the transcription row. No recording covers this provider.","sources":["s1115"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/transcription.rb","method":"transcription_url"}]},"dubbing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no dubbing endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"fast_inference":{"v1":"missing","v2":"native","offered":true,"notes":"GLM-5.3-FlashX is a faster model, documented at 200 tokens per second, and is selected by its model id in the registry. No speed tier request parameter is documented.","sources":["s1098","s1120"],"code":[{"version":"v2","path":"lib/ruby_llm/models.json","method":"glm-5.3-flashx"}]},"file_management":{"v1":"na","v2":"na","offered":false,"notes":"The file API documents upload only, with no list or delete routes.","sources":["s1105","s1106","s1119"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"file_search_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no file search store API.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"fill_in_middle":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no completions endpoint or suffix parameter.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"fine_tuning":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no fine-tuning jobs.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"google_maps":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no maps grounding.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"guardrail_configuration":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no guardrail settings.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"hosted_conversations":{"v1":"missing","v2":"out_of_scope","offered":true,"notes":"The GLM Slide/Poster Agent, a beta, keeps a provider-stored conversation through conversation_id and a history endpoint at /v1/agents/conversation. This is outside RubyLLM coverage; conversation state stays in the application.","sources":["s1121","s1122"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"image_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint. Asynchronous image generation handles one request at a time.","sources":["s1105","s1106","s1112"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"interactions_api":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no such API.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"json_mode":{"v1":"missing","v2":"passthrough","offered":true,"notes":"response_format with type json_object turns on JSON mode, documented for text models. with_provider_options(response_format: {type: \"json_object\"}) forwards it. with_schema is the shared schema API.","sources":["s1097","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/chat.rb","method":"with_provider_options"},{"version":"v2","path":"lib/ruby_llm/protocol.rb","method":"deep_merge"}]},"mantle_protocols":{"v1":"na","v2":"na","offered":false,"notes":"This row is a Bedrock route, and the inventory has no equivalent.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"manual_compaction":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no compaction endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"music_generation":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no music generation.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"nonchat_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"omni_video_generation":{"v1":"na","v2":"na","offered":false,"notes":"The video endpoint takes a prompt and first and last frame images, with no audio or video references.","sources":["s1105","s1106","s1114"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"partner_model_protocols":{"v1":"na","v2":"na","offered":false,"notes":"The inference routes document GLM models and no partner model protocols. The Anthropic Messages endpoint is a compatible route for GLM models.","sources":["s1105","s1106","s1123"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"prefix_completion":{"v1":"na","v2":"na","offered":false,"notes":"The chat schema has no prefix option, and the inventory documents no completions endpoint.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"programmatic_tool_calling":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no programmatic tool calling.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"prompt_cache_controls":{"v1":"na","v2":"na","offered":false,"notes":"Caching is implicit and the chat schema has no cache key or lifetime option. RubyLLM sends no cache parameters on this provider, even with with_caching(key:).","sources":["s1105","s1106","s1109","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"provider_server_fallback":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no provider-managed fallback.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"realtime":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no realtime session API. Streaming transcription uses server-sent events.","sources":["s1105","s1106","s1115"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"response_caching":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no generation response cache. The no_cache option on Web Reader controls page fetching only.","sources":["s1105","s1106","s1107"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"responses":{"v1":"missing","v2":"missing","offered":true,"notes":"The API key guide and Coding Plan quick start list an OpenAI Responses endpoint at api.z.ai/api/v1. No API reference describes it, so its fields are undocumented. RubyLLM declares only Chat Completions for Z.ai, so protocol: :responses raises an error.","sources":["s1123","s1124"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"},{"version":"v2","path":"lib/ruby_llm/provider.rb","method":"fetch_protocol"}]},"responses_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"search_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /paas/v4/web_search returns structured search results, separate from chat. RubyLLM has no search operation for it.","sources":["s1125","s1101"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_apply_patch":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no apply patch tool.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_tool_image_generation":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no image generation tool.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_tool_search":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no tool search.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"server_x_search":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no X search tool.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"shell_computer_tools":{"v1":"na","v2":"na","offered":false,"notes":"The chat tool schema has no shell or computer tools.","sources":["s1105","s1106","s1095"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"sound_effects":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no sound effects generation.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"streaming_speech":{"v1":"na","v2":"na","offered":false,"notes":"The SDK pages list this as a feature, but the docs index, OpenAPI file and pricing page show no endpoint or model for it, and GET /models lists no model for it. The live route answered 400 Unknown Model for the ids tried. A block on RubyLLM.speak would stream from audio/speech through the inherited protocol.","sources":["s1104","s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/protocols/chat_completions/speech.rb","method":"stream_speech"}],"verification":[{"status":"unavailable","checked_on":"2026-11-09","model":"cogtts","notes":"POST /audio/speech returned 400 code 1211: Unknown Model."}]},"task_budgets":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no task budget option.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"text_intelligence":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no text analysis endpoint. The translation agent is a separate agent operation.","sources":["s1105","s1106","s1121"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"vector_store_management":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no vector store or knowledge base API.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"video_batches":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no batch endpoint.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"video_editing":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no video editing input on the video endpoint.","sources":["s1105","s1106","s1114"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"video_extension":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no video extension option.","sources":["s1105","s1106","s1114"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"voice_cloning":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no voice cloning.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"voice_conversion":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no voice conversion.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"web_fetch_api":{"v1":"missing","v2":"missing","offered":true,"notes":"POST /paas/v4/reader fetches a URL and returns parsed content in selectable formats. RubyLLM has no operation for it.","sources":["s1107"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]},"websockets":{"v1":"na","v2":"na","offered":false,"notes":"The inventory documents no WebSocket transport.","sources":["s1105","s1106"],"code":[{"version":"v2","path":"lib/ruby_llm/providers/zai.rb","method":"protocol"}]}},"gaps":[{"notes":"Responses API: the listed api.z.ai/api/v1 endpoint cannot be selected, because RubyLLM declares only Chat Completions here.","sources":["s1123","s1124"]},{"notes":"Standalone search API, Standalone web fetch API: Web Search and Web Reader have no RubyLLM operations.","sources":["s1125","s1107"]},{"notes":"Image generation, Transcription, Streaming transcription: these use inherited Chat Completions code with no recording for this provider. GLM-Image and GLM-ASR-2512 are not in the registry and need assume_model_exists: true. Paint also sends n, and a null size when none is given, which the schema does not describe.","sources":["s1112","s1115"]},{"notes":"Tool choice: Z.ai documents only auto, so other choices are left for the API to answer.","sources":["s1102"]},{"notes":"Thinking controls: GLM-5.3 and GLM-5.3-Flash return an error if thinking is disabled and take only low, high and max effort.","sources":["s1099"]}]}]}
