RAG has two steps. First a search runs over your own material — contracts, manuals, minutes — and picks the few paragraphs that fit the question. Then the model receives those few paragraphs alongside the question, and composes its answer out of them. You do not retrain the model itself, and nothing goes into it permanently.
The effect is large and measurable. In a 910-question measurement the same model reached 48.1% accuracy in its base state, 50.4% after fine-tuning, and 87.5% with RAG. For new knowledge it is therefore not one option among several: it is an order of magnitude better than fine-tuning — which on another model made things distinctly worse.
The most common misconception is that RAG puts an end to hallucination. It does not: it reduces it, because the model has something to lean on, but if the search brings back the wrong passage the answer will be confidently wrong. The larger part of the quality rides on the search, not on the model.
How a working RAG is put together, and where it tends to break, is the subject of Teaching AI Your Own Documents: RAG Explained, a post of its own. Directly related terms: chunking, embedding and hybrid search.