Purpose/Significance To explore a medical knowledge-driven hybrid retrieval-augmented generation method (Med-HyRAG), so as to address the issues of insufficient factual accuracy and low reliability of large language model (LLM) in the medical vertical domain. Method/Process A medical knowledge base focused on rare diseases is constructed. The keyword matching-based sparse retrieval and semantic similarity-based dense vector retrieval are comprehensively applied, and the reciprocal ranking fusion algorithm is employed to optimize and rank the dual retrieval results, thereby providing an optimal context for LLM through hybrid retrieval-augmented generation. Result/Conclusion The indicators such as accuracy, recall of this method are superior to those of single retrieval and other retrieval strategies, and it has good robustness in the comparison of several LLMs. This study contributes to advancing the intelligence of medical knowledge services.