Why AI makes things up about your business, and how RAG fixes it

The model has never read your documents. Its knowledge comes from training, which means, broadly, the public internet up to a cutoff date. Your contracts, prices, policies and decisions were never in it. The danger is that it answers questions about them anyway, fluently: ask it about the world and it is brilliant, ask it about your business and it invents.

The fix has a three-letter name. RAG, retrieval-augmented generation, changes what the model is asked to do. Instead of answering from memory, the system first searches your documents, pulls the most relevant passages, and hands them to the model together with your question. The model then answers from what it was just shown, not from what it half-remembers. The term comes from a 2020 paper by Patrick Lewis and colleagues at Meta's AI research group (arXiv:2005.11401).

The clearest way to hold the difference is an exam. A plain model takes a closed-book exam: it answers from memory and bluffs where memory fails. RAG makes the exam open book, with the relevant pages on the desk. This is why two AI tools built on the same model can behave completely differently. One is reciting. The other is reading.

Underneath, it is three steps, one pattern. Retrieve: search the document library for the passages most relevant to the question. Augment: put those passages into the prompt, next to the question. Generate: the model writes an answer grounded in what it was just handed, and can cite it. Almost every serious "chat with your data" product, from enterprise copilots to internal knowledge assistants, is this pattern underneath.

One honest caveat: RAG reduces invention, it does not eliminate it. The model can still misread or overstate what a passage says. Which leads to the two questions worth asking of any AI tool. Where did this answer come from? Grounded systems can show their sources, and if a tool cannot point to the passage behind its claim, treat the claim as a guess. And what is in the library it reads from? RAG is only as good as what it retrieves, and outdated or wrong documents produce confident, well-cited, wrong answers.

The model was never the database. Its memory is general, frozen, and blurry at the edges. Your advantage lives in your own documents, and the winning setup is not a smarter model. It is your knowledge, wired into the model's reasoning.