Large language models decide whether to retrieve external knowledge based on the semantic relevance of the current prompt, similar to the recall rate problem in RAG (Retrieval-Augmented Generation)