Retrieval-Augmented Generation (RAG) is a pattern where a system retrieves relevant documents from your data before a language model generates a response. Instead of relying solely on what the model learned during training, it pulls from your content at query time.

LLMs don't know your proprietary data, and they hallucinate. RAG addresses both problems at once. It's the most practical path to a model that knows your business without retraining or a fine-tuning cycle that takes weeks and a significant budget.

The catch: retrieval quality is the actual limiting factor. If your documents are poorly chunked, badly indexed, or irrelevant to the query, the model confidently answers with garbage. Most failed RAG implementations fail at the retrieval step, not the generation step. Getting this right means investing in embeddings, chunk strategy, and reranking — not just wiring up a vector database and calling it done.

The pattern was formalized in a 2020 Meta AI paper that framed retrieval as a way to ground generation in external knowledge. It has since become the default architecture for knowledge-intensive applications: internal search, customer support, document Q&A. For most teams, it's also the right first move before considering fine-tuning.

RAG doesn't eliminate hallucination — it reduces it. Keep evals running in production so you know when it drifts.