In real-time session systems supported by the Retrieval Augmented Generation framework, improving retrieval efficiency proves crucial for enhancing overall system performance. Within professional domains, several primary challenges emerge due to extensive external knowledge base and rigorous knowledge segmentation: data noise contamination, prolonged retrieval times, and reduced accuracy in recall results. To address these issues, we propose a novel RAG framework called Focused Retrieval Augmented Generation (FRAG), which aims to greatly narrow down the retrieval scope on the external knowledge base. Our framework constructs a retrieval mapping table on the external knowledge base and aligns it with the vector database, and applies the mapping structure to the retrieval process. For experimental validation, we construct an extensive external knowledge base and corresponding question-answering dataset using university campus documentation. Experiment results demonstrate that the FRAG framework successfully narrows the retrieval scope to 1/56 of its original size, substantially enhancing retrieval performance through more accurately focused retrieval scope.