Command R-plus RAG 기반 한의진단학 소규모 언어모델(SLM) 개발 및 성능 평가

Abstract

This study focuses on the utility of Small Language Models (SLMs) optimized for specific domains, aiming to develop an artificial intelligence model specialized in the field of Korean Medicine diagnostics. For this purpose, we utilized a quantized version of the state-of-the-art SLM, Command R-plus, in a local environment and integrated it with Retrieval-Augmented Generation (RAG) technology. The knowledge base for the RAG system was constructed using textbooks from Korean Medicine colleges, specifically ‘The Diagnostics in Korean Medicine: Examination’ and ‘The Diagnostics in Korean Medicine: Diagnosis’, utilizing a FAISS vector store and a multilingual embedding model. The model’s performance was evaluated using a dataset of 70 questions related to Korean Medicine diagnostics, and was compared with the base model (without RAG) and a large language model, ChatGPT-5. The results showed that the RAG-applied model achieved an accuracy of 70.00%, a considerable improvement over the base model’s 48.57%. While not matching the 81.43% accuracy of ChatGPT-5, it significantly closed the performance gap and demonstrated strong potential. This research demonstrates that combining SLM with RAG technology can effectively process specialized knowledge in a niche field like Korean Medicine, presenting its potential for future development as an educational and clinical support system.