Field Context
Most hallucination complaints in enterprise assistants are retrieval failures in disguise. The model responds confidently because context appears complete, but the retrieval layer selected stale or low-signal chunks.
What We Changed
We split pipeline ownership into ingestion quality, retrieval ranking, and response policy. Each stage has its own checks and rollback route, which makes drift visible before it impacts users.
Practical Execution Notes
Document freshness, source trust, and confidence thresholds should be treated as first-class production settings with clear operators.
- Store chunk version metadata so bad imports can be rolled back quickly.
- Log retrieval misses and low-confidence answers as operational incidents.
- Route ambiguous prompts to human escalation with source trace attached.
Where Teams Usually Slip
Indexing every available file feels safe but usually reduces precision. Retrieval quality improves when you intentionally narrow what can be cited for each query type.
Reliable RAG is built with governance and observability, not only embeddings.