<p>In the domain of medical question answering, ensuring high accuracy, reliability, and explainability is paramount. While Large Language Models demonstrate immense potential, their inherent risk of hallucination precludes direct application in clinical settings. Augmenting these models with external knowledge graphs through Retrieval-Augmented Generation has thus become a prevailing paradigm. However, existing methods are often impeded by the retrieval of redundant evidence and a lack of diversity. To address these challenges, we propose RaR, a novel retrieve-and-rerank framework. RaR first employs a hypothesis-driven exploration mechanism that leverages a model’s prior knowledge to simplify knowledge graph reasoning and efficiently retrieve evidence. Subsequently, it introduces a pioneering clustering-based reranking mechanism to simultaneously eliminate redundancy and maximize the diversity of reasoning paths, providing high-quality, multi-perspective knowledge support. To our knowledge, RaR is the first work to apply clustering for reasoning path selection in medical dialogue scenarios. We conducted extensive experiments on four publicly available English medical long-text dialogue datasets. The results demonstrate that, compared to multiple mainstream LLMs baselines (such as GPT-4 and Llama), RaR significantly improves performance across all evaluation benchmarks and enhances the interpretability of model responses.</p>