Retrieval-Augmented Generation (RAG)

Status: public · Confidence: medium (0.815) · Basis: verified_sources

## TL;DR

Retrieval-Augmented Generation, or RAG, is a language-model architecture pattern that retrieves external evidence and conditions generation on that evidence. It is useful when an answer should be grounded in a document collection rather than only in a model's stored parameters.

## Core Claims

The original RAG formulation combines a generator with a retriever over external text. The retriever selects documents or passages; the generator uses retrieved evidence while producing an answer.

Dense Passage Retrieval is a common retrieval foundation for this pattern. It represents questions and passages as dense vectors, enabling nearest-neighbor retrieval for open-domain question answering.

REALM is an adjacent retrieval-augmented pretraining approach: the model learns to retrieve text from a corpus and use that retrieved text during language-model pretraining.

## Citation Boundaries

Use this article for stable RAG concepts: retrieval, external memory, dense passage retrieval, and evidence-conditioned generation. Do not use it to claim a system is factual unless the specific retrieval corpus, ranking, and citation behavior have been evaluated.

## Further Reading

- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401)
- [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906)
- [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909)