01 Personal · open source
Relay: AI support triage engine
Here the LLM decides which action to take instead of just generating
text. It can answer, search the knowledge base, ask for clarification, escalate to a human
or block the input. Each action has its own rules: respond_kb only runs on
high confidence, and too many clarifications trigger an automatic escalation.
Semantic search runs on pgvector, with a lexical fallback and ranking based on answers that already worked. I moved embedding generation out of the request into an async worker, which cut chat responses from 2 to 3 seconds down to about 800 ms. It has human review, Prometheus metrics and tracing.
- Agent actions
- respond
- respond_kb
- clarify
- escalate_ticket
- resolve
- block