RAG (retrieval-augmented generation) is a technique in which a language model, before answering, retrieves information from a connected knowledge base – documents, a price list, a database – and answers based on those sources rather than its training memory alone.
The mechanics in three steps: the user asks a question → the system finds matching fragments in your sources → the model composes an answer from those fragments, often citing them. The model doesn't have to "remember" your offer; it gets it served at the moment of the question.
A practical example: an assistant for an accounting firm's clients answers from the firm's current price list and procedures – not from whatever the model "heard" about accounting firms on the internet. When the price list changes, you swap the document.
Why it matters for business: RAG is the core of most sensible AI implementations in companies. It lets the model work on your own, current data without costly training, and it clearly reduces hallucinations, because answers have sources. This is how you build AI that knows your company.