An open-book approach
Imagine asking someone a question about a long handbook. Instead of expecting them to remember every page, you give them the relevant sections before they answer. Retrieval-augmented generation applies a similar idea to a language model.
Two jobs, one answer
The retrieval step finds candidate material. The generation step uses that material to compose a response. Each can fail independently. A fluent answer is not proof that the search found the right passage, or that the answer represented it faithfully.
Good documents still matter
Clear headings, current versions, and manageable sections help make a collection useful. Conflicting documents and stale instructions can confuse a retrieval system just as they confuse people. Adding AI does not remove the need to maintain the library.
Make checking easy
A useful response links back to the passages it relied on. Those references should actually support the claims beside them. If the material does not contain an answer, the system needs a way to say so instead of filling the gap with a plausible guess.
The takeaway
Retrieval is a way to supply context, not a guarantee of correctness. Treat the source collection, the search step, and the generated answer as separate things to inspect.
