Language models can look things up instead of memorizing everything. A 2020 approach pairs a generator with a retriever that fetches relevant documents at query time. It grounds answers in external knowledge and eases the problem of fabricated facts.
Fetch then generate
The pipeline has two stages. A retriever finds passages, and the model conditions its answer on them. Knowledge lives outside the weights.
Fresh knowledge
Updates are easy. Swapping the document store refreshes what the system knows without retraining. It stays current.
Fewer fabrications
Grounding helps. Answers tied to sources are less prone to invention. Reliability improves.
Attribution
Sources can be shown. The retrieved passages provide citations for a claim. Verification becomes possible.
Enterprise appeal
Business adopted it. Companies connect models to private documents securely. It became a dominant deployment pattern.
Open challenges
Retrieval is imperfect. Poor passages still mislead the generator. Improving relevance remains active work.
The bottom line
Retrieval-augmented generation grounds models in external documents fetched at query time. It refreshes knowledge and curbs fabrication. The pattern now dominates practical deployments.