Retrieval-augmented generation grounds a model's answer in passages fetched from an external index at query time. This alone cuts hallucination rates by roughly half compared to a model answering from parametric memory only. The retrieved passages are typically ranked by a dense encoder trained jointly with the generator. Most production pipelines refresh the index nightly rather than serving it fully live. Chunk sizes between 200 and 500 tokens tend to balance recall against context dilution.