Retrieval-Augmented Generation (RAG)
A technique where an AI model looks up relevant external documents at answer time instead of relying only on training data.
What is Retrieval-Augmented Generation (RAG)?
Retrieval-Augmented Generation (RAG) is a technique that lets an AI model consult external documents at the moment it's answering a question, rather than relying purely on what it absorbed during training. Instead of a chatbot answering entirely from memorized patterns, a RAG system first searches a knowledge base or the web for relevant passages, then writes its answer grounded in what it just retrieved, similar to a person looking something up before answering rather than answering from memory alone.
Why it matters
The technique was formalized in the 2020 paper "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" by Lewis et al., presented at NeurIPS 2020, which paired a pretrained sequence-to-sequence language model with a dense vector retriever pulling passages from Wikipedia and showed the combination produced more specific, diverse, and factually accurate output than a generation-only baseline on open-domain question answering. Mechanically, most RAG systems convert a query into a vector embedding, search a vector database for the most similar stored document chunks, and insert those chunks into the model's context window alongside the original question before generation happens. RAG is now the standard architecture behind AI systems that need to cite current, source-specific, or proprietary information (including AI shopping assistants pulling from product feeds and structured data rather than a static training snapshot) precisely because it reduces hallucination and lets an answer be traced back to a specific retrieved source.