The core idea

Look it up, then answer.

RAG retrieves relevant text from your own data and puts it in the prompt, so the model answers from real sources instead of memory alone.