Retrieval-augmented generation, or RAG, pairs a language model with a search step. When a question arrives, the system first retrieves the most relevant chunks of your own content, then hands those chunks to the model as context before it writes an answer. The model effectively reads your material fresh for every request, so it never has to guess when the answer sits right in front of it.
A support chatbot that cites your help docs is the classic example. Instead of hoping the model memorized your refund policy, you store your documentation in a searchable index, pull the passages that match the customer's question, and let the model summarize them. Answers stay current because you update the documents, not the model.
This solves two problems at once. It keeps responses grounded in facts you control, which cuts down on invented answers, and it lets the model work with private or recent information that was never in its training data. For most startups, RAG is far cheaper and faster to ship than fine-tuning.
The quality of a RAG system lives almost entirely in the retrieval step. If search returns the wrong passages, even a strong model gives a wrong answer. Good chunking, a solid vector database, and sensible ranking matter more than the specific model you pick. Logging which passages were retrieved also makes failures easy to debug.