Retrieval-augmented generation searches a prepared knowledge source for relevant evidence, places that evidence in context, and asks the model to answer from it.
A researcher receives a question, finds the most useful pages in a library, and keeps those pages open while composing an answer.
A retriever selects relevant text chunks. The model generates from the question plus those supplied passages.
Retrieval happens before generation. It changes the model’s working material; it does not rewrite the model itself.
Galleries close at 5:00 PM Tuesday–Sunday.
Special closures supersede normal hours.
Embeddings help rank semantic similarity. Metadata filters can narrow the candidates by source, date, category, or access policy.
Poor chunks or vocabulary mismatch can hide the right passage.
Retrieving outdated material grounds the answer in outdated material.
The model can still overgeneralize beyond the supplied evidence.
Preserve source identity, return citations when possible, separate retrieved text from instructions, enforce access controls before search, and say when the evidence does not answer the question.
Chunk and organize authoritative sources.
Find the few passages most relevant now.
Use the evidence inside the model’s context.