Development of a Meta-Question Enhanced Retrieval-Augmented Generation Model and Its Application in Dermatology

RAG (Retrieval Augmented Generation) provides a solution for enhancing domain-specific knowledge in large language models by interacting with external information. RAG functions similarly to a search engine, retrieving the most relevant knowledge or related conversation history to the user's query. It then combines this with the original query to create a detailed prompt, guiding the model to generate accurate outputs. However, traditional RAG methods match short queries with long texts in the knowledge base, leading to mismatched results despite apparent similarities due to length discrepancies. To address this issue, we propose the MQ-RAG (Meta Questions Retrieval Augmented Generation) method, a question-answer pair-based approach built upon RAG. The key difference in MQ-RAG lies in the retrieval and recall content: it uses a database of Meta Questions, which are pairs of questions and corresponding answers. When using MQ-RAG, queries are matched with Meta Questions to retrieve the most relevant answers, ensuring high compatibility between the query and the retrieved content. For evaluation, 100 multiple-choice questions from the Dermatology and Venereology Medical Licensing Examination database were selected. Results showed that, using Qwen2-7b and Qwen2-72b as base models, MQ-RAG improved accuracy by 17.14% and 2.27%, respectively, compared to traditional RAG. Our MQ-RAG method demonstrated higher accuracy across both large and small parameter models, with particularly significant improvements in smaller models.