The core idea of RAG is to combine the language capabilities of a general-purpose large model with the specialized content of an enterprise's private knowledge base, first retrieving relevant document segments and then feeding them as context to the large model to generate answers.