Boring part
End to End flow
- User sends a query -> Orchestrator receives via API
- Input guardrails validate and sanitize the query
- Receiver searches knowledge base (BM25 + Vector) and returns Top-K chunks
- Augmented prompt is built using query + retrieved context + history + instructions
- LLM Generate a response
- Output guardrails validate, redact and ensure compliance
- Response is returned to the user via FastAPI
