AI-generated Answer for celestial_beagle_61127
Short answer: the platform does not expose the raw, fully‑assembled LLM prompt per turn. What you can log is (a) all turn‑by‑turn user/agent messages and tool calls in real time, and (b) the RAG context that was retrieved for each turn via conversation history/webhooks. If you need the exact prompt payload that the LLM received, use a Custom LLM and log the incoming chat‑completion request on your server.
What to log and where to get it
- Live, per turn:
- Subscribe to the WebSocket events and log:
- user_transcript (finalized ASR text)
- agent_response (final agent message)
- client_tool_call and agent_tool_response (function/tool use)
- contextual_update you send (UI or app context; these are recorded in history as tool calls) (elevenlabs.io)
- After the call (for RAG/debugging):
- Get conversation history: GET /v1/convai/conversations/:conversation_id returns transcript, metadata, analysis, and initiation client data; the endpoint was updated to include RAG information for retrieved KB content. (elevenlabs.io)
- The dashboard history view also shows “retrieved chunks by RAG” per call (useful for correlating what context was fed to the LLM). (elevenlabs.io)
- Or enable post_call_transcription webhooks to receive the same conversation object automatically when analysis completes (same structure as GET Conversation). (elevenlabs.io)
About logging “the prompt”
- There is no documented API/setting to return the exact compiled LLM prompt per turn (system prompt + dynamic vars + RAG context + tool traces). This is not shown in the Events or Conversations schemas. (elevenlabs.io)
- If you must see the exact payload the LLM receives, integrate a Custom LLM (OpenAI‑compatible chat completions). Your server will receive the full messages/tools payload for each turn, which you can log before responding. (elevenlabs.io)
References
- Agent WebSockets (event types and examples): user_transcript, agent_response, client_tool_call, contextual_update. (elevenlabs.io)
- Conversations API: list/get conversation; GET was updated to include RAG info. (elevenlabs.io)
- Changelog: RAG chunks visible in history view. (elevenlabs.io)
- Post‑call webhooks: delivers the conversation (same model as GET Conversation). (elevenlabs.io)
- Custom LLM integration (log the incoming request to capture the exact prompt/context). (elevenlabs.io)
If you want, I can sketch a minimal logger that:
- writes WebSocket events to disk during the call, and
- upserts the final GET Conversation (including RAG fields) when the call ends. (elevenlabs.io
crystal hamlet
