The core idea of RAG is to have the AI first retrieve relevant information and then generate an answer based on the retrieved context, rather than answering from memory alone.