Giving a model knowledge it never learned (RAG).

RAG doesn't teach the model a thing. The weights never change — the knowledge is borrowed for one call, pasted in, then gone. Open-book, every single question.

The one idea

RAG (retrieval-augmented generation) gives a model knowledge without retraining: retrieve the relevant text at query time and paste it into the context, so it answers open-book from borrowed, current, specific text.

A model's knowledge is frozen at training, and it's generic. It doesn't know your company's docs, today's news, or this user's data. You could retrain it — expensive and slow. Or you could just hand it the right pages at the moment of the question. That second trick is RAG, and it quietly changes what a model can do. RAG stands for retrieval-augmented generation, and the idea is open-book. When a question comes in, you first search a knowledge source — your documents, a database — for the passages most relevant to it. Then you paste those passages into the context, right next to the question, before the model ever answers. Retrieve, then generate. Now the model answers using that supplied text. It isn't recalling from memory; it's reading the page you just handed it and reasoning over it. The knowledge is borrowed for this one call — pulled in fresh, used, then gone.

How it works — the demo

A question arrives; the system retrieves the most relevant passages from your docs and pastes them into the context; the model answers from that supplied text, not its frozen memory.

Next question, you retrieve again. The model supplies the reasoning; you supply the facts. This quietly solves three problems at once. The cutoff: retrieve today's information, and the answer is current. Private data: point it at your own documents, and it knows them. And hallucination drops, because the answer is grounded in real retrieved text instead of guessed from frozen training. Fresh, private, and grounded — from one mechanism. And notice what this is: the prompt is the program, in action. You're engineering the context by inserting the exact knowledge the model needs, automatically, per question. The model becomes a reasoning engine over facts you supply — not a fixed memory bank you cross your fingers actually contains the answer. Retrieval writes the program. The trap: thinking RAG teaches or trains the model.

The trap to avoid

Thinking RAG "teaches" or "trains" the model — it never touches the weights; the knowledge is borrowed for one call.

Why it matters — and what’s next

It all hinges on finding the RIGHT passages — how does that search work?

It doesn't touch the weights at all. Nothing is learned. You're placing text in the context at query time, for one answer. That's why it's so flexible: change the documents on the shelf, and the very next answer changes — no retraining, no waiting, no cost of touching the model. So RAG gives a model knowledge it never learned — retrieve the right text, paste it in, answer. Borrowed, current, specific, grounded. But the whole thing hinges on one hard step: finding the right passages out of millions. Get that wrong and everything downstream fails. So how does that search actually work? Next: how semantic search finds things.

This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.