Retrieval-augmented generation
Also called RAG.
What Retrieval-augmented generation means
Retrieval-augmented generation is a technique in which a language model fetches documents at question time and conditions its answer on them, instead of relying only on what is held in its weights. The name comes from a 2020 paper by Patrick Lewis and colleagues, presented at NeurIPS 2020.
More precisely
The paper describes models that "combine pre-trained parametric and non-parametric memory for language generation", where "the parametric memory is a pre-trained seq2seq model and the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever." Deployed systems swap that fixed Wikipedia index for a live web index, a product catalogue or a single site's own pages, but the shape is unchanged: retrieve first, then generate conditioned on what came back.
Why it matters
This is the mechanism that lets a page published last week be quoted by a model trained long before it existed. Being retrievable is a different problem from being in the training data, and it is the one a publisher can actually act on.
Example
An assistant asked about a local business searches, pulls a handful of pages, and writes its answer from those pages rather than from memory.
Where this came from
Definitions here follow the specification or documentation that defines the term, not a paraphrase of another glossary.
-
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Lewis et al., NeurIPS 2020
The origin of the term, the authorship and 22 May 2020 submission, and the definition combining parametric seq2seq memory with a non-parametric dense vector index accessed by a neural retriever.
Primary source · read 2026-09-03