Write apps/ai_agent/docs/concepts-prompts-and-models.md: the system prompts for each LLM-backed endpoint, the models used, and how outputs are parsed and validated back into typed responses.
Acceptance criteria:
- Each prompt reproduced verbatim with the behaviour it is designed to produce
- Documents the model used per endpoint and how to change it safely
- Documents output parsing and the fallback path when the model returns malformed or partial JSON — the transfer analyser's default handling of a missing
confidence or flagged is the worked example
- States the privacy boundary explicitly: which data is sent to the model provider and which never leaves the system
Write
apps/ai_agent/docs/concepts-prompts-and-models.md: the system prompts for each LLM-backed endpoint, the models used, and how outputs are parsed and validated back into typed responses.Acceptance criteria:
confidenceorflaggedis the worked example