Definition
A technique where an AI retrieves relevant documents from a knowledge base before generating a response. Improves accuracy on factual queries and reduces hallucination. Instead of relying only on training data, the system looks up the relevant passages first and writes its answer from those, which is how it can quote a business's own prices and policies.
What it looks like in practice
A caller asks an HVAC company's after-hours line whether a mini-split install includes the electrical work. A general AI would guess; most installs do, or check with your contractor. A retrieval-based one searches the company's own uploaded documents first, finds the install scope sheet, and answers what that sheet actually says: line-set and condensate are included, a new dedicated circuit is quoted separately by the electrician, and the flat diagnostic fee applies if the panel needs evaluation. The caller gets a real answer on a Sunday. When the owner raises that fee in March, he edits the sheet, and the answer changes with it.
Why it matters
RAG is why an assistant that has read your price sheet answers differently from one that has not. The first gives a caller the actual number and what it covers; the second says 'contact us for details', and the caller who wanted an answer goes and asks the next company. So the real work is not the AI, it is the source material, which documents it can see, and who keeps them current. Accuracy is inherited. An out-of-date policy quietly becomes an out-of-date answer delivered with total confidence, which is a hallucination with a paper trail behind it.