TCM MLKG-RAG: Traditional Chinese Medicine Intelligent Diagnosis Based on Multi-Layer Knowledge Graph Retrieval-Augmented Generation

Traditional Chinese Medicine (TCM) search engines often struggle with the issue of redundant data volumes, making it difficult to meet users' demands for precise information retrieval. Large Language Models (LLMs) excel in understanding questions and summarizing key points due to their vast number of parameters. However, keeping pace with updates in TCM knowledge requires significant computational resources and time for finetuning these large models. Retrieval-augmented generation (RAG) allows LLMs to generate more accurate, specialized, and timely responses without the need to update their parameters. TCM knowledge is characterized by its dispersed nature and a blend of classical and vernacular language, which makes traditional RAG unsuitable for the field of TCM. To address this, we have developed a TCM knowledge graph RAG that integrates multi-layered knowledge bases, with the lower layer consisting of TCM-specific terminology and explanations, and the upper layer comprising clinical diagnosis and treatment cases. Furthermore, we have proposed two retrieval methods: keyword retrieval and therapy retrieval. Keyword retrieval is designed to search for information on TCM-specific terms, while therapy retrieval locates diseases based on medical and patient information and provides corresponding treatment methods. We have validated the effectiveness of our methods across various datasets.