Adapting Large Language Models for Healthcare with an Enhanced Retrieval-Augmented Generation Framework

In recent years, the application of large language models (LLMs) in the healthcare domain has garnered significant attention. However, the specialized nature of medical knowledge poses considerable challenges for LLMs in handling knowledge-intensive tasks. Traditional fine-tuning paradigms, which rely on static pretraining data, struggle to integrate the latest medical research and real-time clinical guidelines. Furthermore, LLMs may be prone to generating hallucinations when providing medical suggestions, limiting their practical utility. Retrieval Augmented Generation (RAG) partially mitigates the constraints of static knowledge and hallucinations by incorporating external knowle-dge bases but remains vulnerable to interference from irrelevant documents. To address these issues, we introduce RAFT4MED, a retrieval augmented generation framework specifically designed for medical question-answering. RAFT4MED leverages medical knowledge base retrieval enhancement with supervised fine-tuning, explicitly modeling the reasoning process from question understanding and knowledge retrieval. Comprehensive experiments demonstrate that RAFT4MED outperforms conventional RAG and fine-tuning met-hods on medical benchmark evaluations.