RAG stands for Retrieval Augmented Generation, three words for one honest idea: stop trusting the model's memory. Instead of asking the model to recall facts from training, you fetch your real documents first and hand them over together with the question. The model answers from what is on the table in front of it. An open-book exam instead of a memory test.
This is the pattern behind almost every serious chat-with-your-documents tool, and it earns its place for one reason: hallucinations drop when the model is grounded in your actual material. The model is still predicting text, but now it is predicting from your handbook, your contracts, your knowledge base, instead of a blurry memory of the internet. It can quote instead of invent.
For an operator, the key insight is that a RAG system is only as good as its retrieval. If the fetch step brings back the wrong pages, the model gives a fluent answer grounded in the wrong material, which is worse than an obvious failure because it looks right. When a RAG tool disappoints, the model is rarely the problem. Look first at what was retrieved, and at whether the documents themselves are current and complete.
A concrete example. A company builds an internal assistant for HR questions. Without RAG, the model answers from general internet knowledge of what policies tend to say, which is a hallucination risk wearing a friendly tone. With RAG, the assistant fetches the relevant sections of the actual employee handbook and answers from those, showing where each answer came from. Same interface, same model. The difference is that one build can point at its evidence.
“Hallucinations drop when we RAG it against our own documents.”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.