Common failure modes and fixes
Hallucinated facts
→ strict grounding instruction + abstain-on-low-coverage + show citations
Right answer not retrieved (low recall)
→ better chunking, hybrid retrieval, higher k, query rewriting
Irrelevant chunks crowding the prompt (low precision)
→ re-ranking, lower k, metadata filtering
Lost in the middle (model ignores mid-prompt context)
→ fewer, better chunks; reorder
Stale answers
→ re-index on source change (content hash)
Ambiguous query
→ query expansion / rewriting before retrieval