Bridging Rationales and Relations: The Graph-Rationale-Guided Retrieval-Augmented Generation in Medical QA

Large language models (LLMs) face challenges of hallucination and knowledge obsolescence in medical question answering. Existing retrieval-augmented generation (RAG) frameworks can improve retrieval reliability through rational guiding; however, their neglect of structured knowledge leads to insufficient relational reasoning. This paper proposes the Graph-Rationale-Guided Retrieval-Augmented Generation (GRAG) framework, which introduces a knowledge graph layer based on Rationality-Guided Retrieval-Augmented Generation (RAG²) to support dynamic graph query expansion, evidence fusion, hallucination detection, and quantized low-rank adaptation (QLoRA). GRAG's core mechanisms include rational generation, entity extraction, graph construction based on Unified Medical Language System (UMLS)-Neo4j, and similarity-driven multi-source evidence fusion. Experiments on medical question answering dataset (MedQA), medical muti-choice question answering (MedMCQA), and a self-constructed RareDisease-MedQuAD subset show that GRAG outperforms baseline models by approximately 10-12% in accuracy, reduces hallucination rate by around 20%, and achieves graph fidelity exceeding 80%. Ablation experiments further confirm that the knowledge graph (KG) and QLoRA modules each contribute approximately 5-8% to performance improvements. Overall, GRAG bridges the gap between rational guidance and structured retrieval, providing a more interpretable and reliable solution for MedQA systems.