OphthaRAG: A Retrieval-Augmented Generation Framework for Enhanced Eye Disease Diagnosis Support and Question Answering

The integration of Large Language Models (LLMs) into healthcare holds transformative potential, yet their application in highly specialized fields like ophthalmology is constrained by hallucination and domain-specific knowledge deficits. To address this, we introduce OphthaRAG, a comprehensive Retrieval-Augmented Generation (RAG) framework engineered specifically for ophthalmological diagnosis support and question answering. Grounded in a meticulously curated knowledge base encompassing the Basic and Clinical Science Course (BCSC), clinical guidelines, and peer-reviewed literature, OphthaRAG employs a robust two-stage retrieval pipeline utilizing bi-encoder and cross-encoder architectures. We rigorously evaluate the framework using open-source LLMs, including Llama-3-70B and Mixtral-8x7B, on standardized benchmarks such as OphthoQuestions. Our experimental results reveal substantial improvements: OphthaRAG boosts diagnostic accuracy by 15% to 25% over zero-shot baselines and drastically mitigates hallucination rates, improving context faithfulness by 43%. By enabling traceable, evidence-based reasoning, OphthaRAG bridges the critical gap between theoretical LLM capabilities and the strict reliability requirements of clinical practice.