A framework that retrieves relevant passages first, then feeds query plus evidence to the model so it answers from real text.
Retrieval-Augmented Generation (RAG) is a framework that fetches relevant evidence before the model writes, then feeds the query plus that evidence into the model in a single context window. First proposed by Lewis et al. in 2020, it routes a query through a vector index, pulls the top passages, and lets the model reason over real text instead of inventing facts.
The retrieve-then-generate order is what makes it robust: factuality is built in because the model works from supplied text; failure is cleaner because if retrieval finds nothing the system can say so early; and it is faster, needing one pass rather than draft-then-fix. We have argued this beats Google's generate-then-ground approach, which bolts citations on after drafting.
RAG underpins how modern assistants stay current, and it leans on embeddings and vector search to supply the "retrieve" step. Amazon's Rufus shopping assistant is a production example built on RAG.
