A model answering from memory is answering from a snapshot that's already stale. Retrieval-augmented generation fixes this by handing the model the actual current source material at the moment you ask, so it's reading instead of recalling.
The loop is three steps: pull the relevant passages, paste them in as context, then ask the question against that context only — not what the model "already knows." When the source is too big to paste whole, "pull the relevant passages" becomes a search problem: chunk it, turn each chunk into a vector capturing its meaning, then at query time find the chunks whose vectors sit closest to the question's. That search decides what gets pasted, not the model.
The discipline is in the last step: a good RAG answer cites the pasted material, not vibes. Run the seed: ask cold, then grounded, and compare which answer you'd trust.