Retrieval-Augmented Generation
Your AI isn't dumb — it's just working from memory. RAG gives it the receipts.
Grounded in the research on retrieval-augmented generation
The core problem
A large language model is trained once, then frozen. Everything it knows is baked into its weights at that cutoff date. Ask it about last quarter's earnings, your private codebase, or a paper published six months ago — it either halts, confabulates, or hedges. The weights can't update in real time. RAG is the fix: instead of retraining, you pull relevant documents at query time and drop them straight into the model's context window before it answers.
Where it started
Facebook AI Research — specifically Patrick Lewis and colleagues — published the RAG paper in 2020. They fine-tuned a BART sequence-to-sequence model to retrieve from a dense Wikipedia index using a separate dense passage retrieval component, then generate answers conditioned on those passages. It outperformed pure parametric models on open-domain QA benchmarks like Natural Questions and TriviaQA. The core insight was that retrieval and generation could be trained jointly end-to-end, not bolted together clumsily after the fact.