Build a grounded answer bank from resolved Intercom conversations

Turns the support replies your team has already written into a reusable, searchable knowledge base. Every time a conversation is resolved in Intercom, the workflow extracts the underlying question and the agent's confirmed answer, then stores them as a clean, deduplicated entry with its embedding so future agents and bots can retrieve grounded answers.

When your best answers are trapped inside one-off Intercom threads and never make it into help docs. Run it to bootstrap a RAG corpus from real resolutions, or keep it running to capture new fixes automatically.

1. 1A schedule fires (e.g. hourly) and lists Intercom conversations resolved since the last run.
2. 2For each conversation, the full transcript is fetched.
3. 3An LLM distills the thread into a normalized question, a canonical answer, and a topic tag, discarding pleasantries and PII.
4. 4The answer text is embedded with an OpenAI embedding model.
5. 5A logic step checks Postgres for a near-duplicate by cosine similarity; close matches update the existing row instead of creating a new one.
6. 6The new or merged entry is upserted into the `answer_bank` vector table with a source conversation ID and timestamp.