RAG reduces hallucinations by adding evidence, but it can still give wrong answers. The system may retrieve irrelevant chunks, miss the right document, or pass outdated information to the model.
In practice
Other failures include duplicate documents, poor metadata, weak chunking, missing permissions, ambiguous user questions, and prompts that let the model answer beyond the evidence.
What to watch
Reliable RAG needs testing. Track answer quality, failed queries, source coverage, citation accuracy, and user feedback instead of assuming retrieval solves everything.