Collaborative Dual-Source Retrieval Augmented Generation for Medical Education Question Answering

Large Language Models (LLMs) demonstrate significant potential in medical education question answering, yet their inherent issues of hallucination and knowledge obsolescence limit their application in high-reliability scenarios. While Retrieval-Augmented Generation (RAG) mitigates these limitations by incorporating external knowledge, it is constrained by coverage gaps inherent in static knowledge bases. Techniques such as iterative retrieval have thus been introduced to acquire deeper information from dynamic sources like the open web through multi-turn interactions. However, existing general-purpose deep retrieval mechanisms often lead to semantic mismatches and noise accumulation when processing specialized medical queries, primarily due to a lack of domain adaptation during query formulation. To address this, this paper proposes a Collaborative Dual-Source Retrieval Augmented Generation framework (CD-RAG). CD-RAG employs task-adaptive query rewriting to tailor requests for heterogeneous retrieval sources. It systematically integrates standardized medical knowledge with dynamic, cutting-edge information by combining local hybrid retrieval with iterative web retrieval enhanced by a reflection mechanism. Furthermore, the framework incorporates a semantic based re-ranking model to unify and denoise heterogeneous evidence, constructing high-quality contextual inputs to enable traceable responses from the LLM. Experimental results on the MedQA and RAGCare-QA datasets indicate that CD-RAG outperforms existing methods, effectively enhancing both the accuracy and timeliness of LLMs in medical assessments.