Retrieval Augmented Generation (RAG) is a framework for augmenting Large Language Modeling (LLM) capabilities by retrieving external knowledge from large databases. This approach effectively solves the common problems of factual accuracy, information obsolescence, and “hallucination” in generative language modeling. RAG generates answers that are more demanding in terms of searcher performance and knowledge base content, to address these two problems this paper proposes a method called Extract-Related-Generate for RAG (ERGRAG), which is different from previous RAG methods in that it first decomposes the question, generates a number of sub-questions, and then improves the model's accuracy in responding to multi-hop questions through the model's answers to the sub-questions. This paper evaluates the performance of the model in three datasets, and the ERGRAG method proposed in this paper shows a large improvement in accuracy, indicating that question decomposition is an effective idea to improve the RAG system's response to multi-hop questions, and provides new ideas for subsequent RAG research.