Retrieval doesn't eliminate the risk of hallucination. It changes what gets hallucinated — and that difference matters more than most deployed systems admit.
"We added RAG, so hallucination is basically handled" is one of the most common — and most dangerous — sentences in applied AI work. Retrieval-augmented generation genuinely helps. It just doesn't do what that sentence implies. Grounding a model in retrieved source material reduces one specific category of hallucination: the case where a model would otherwise have to guess because it simply doesn't have the fact. It does nothing to prevent several other, still very real ways a fluent, confident-sounding answer can be wrong.