Does RAG stop a model from hallucinating?

No. RAG lowers the rate of a model inventing facts wholesale, because it gives the model something real to answer from instead of pulling only from its training weights. That’s a genuine reduction, not a guarantee.

Three things still go wrong. Retrieval can hand back the wrong or partial chunks, and the model will answer faithfully from bad material. That’s a retrieval failure, not a hallucination, even though the output looks the same. The model can also ignore correct context it received, favoring what it learned in training over what’s sitting right in front of it in the prompt. And the corpus itself can be stale or self-contradictory, in which case the model faithfully repeats whatever version of the truth it was handed, right or wrong.

RAG changes what the model has to work with. It doesn’t change whether the model uses it correctly.