BiPharm-RAG: A Dual Hypergraph Retrieval-Augmented Large Language Model for Traditional Chinese Medicine Reasoning

BiPharm-RAG is a retrieval-augmented large language model system designed for the traditional Chinese medicine (TCM) domain. Its core innovation is a cross-source dual hypergraph retrieval architecture, which integrates large language models with TCM knowledge to enable end-to-end intelligent diagnosis and treatment reasoning from symptom analysis to prescription recommendation. It also supports knowledge provenance to ensure that recommendations are evidence-based.

Traditional Chinese medicine knowledge is scattered across ancient books, medical records, and other carriers, making it difficult to reuse quickly. Large language models (LLMs) are prone to the "hallucination" problem in the TCM domain. This project aims to combine the reasoning capabilities of LLMs with structured TCM knowledge to build a trustworthy intelligent diagnosis and treatment system.

It adopts a two-layer hypergraph structure:

1. Symptom-syndrome hypergraph: Associates multiple symptoms with syndromes through hyperedges to capture the holistic view of syndrome differentiation;
2. Syndrome-prescription hypergraph: Connects syndromes, treatment methods, prescriptions, and medicinal materials to support end-to-end reasoning; Cross-source integration of classic ancient books, modern medical records, medicinal material databases, and syndrome standards, realizing knowledge integration through node alignment.

1. Symptom semantic encoding: Understand synonymous expressions, degree descriptions, and implicit associations;
2. Hypergraph retrieval: Recall relevant syndrome candidates;
3. Syndrome screening: Rank based on comprehensive matching degree, frequency, and context;
4. Prescription generation: Retrieve classic prescriptions based on syndromes and generate personalized recommendations;
5. Knowledge provenance: Attach authoritative literature references for easy review.

• Hypergraph Convolutional Network (HGCN): Captures high-order associations in TCM knowledge;
• Contrastive learning: Improves retrieval accuracy;
• Multi-task joint training: Synchronously optimizes symptom encoding, syndrome classification, and prescription generation.

• Clinical assistance: Provides quick retrieval and reference for physicians;
• TCM education: Simulates diagnosis and treatment scenarios to assist students in practice;
• Ancient book mining: Discovers implicit knowledge associations;
• Primary healthcare: Provides TCM service support for general practitioners.

Current limitations: Dependence on data quality, insufficient support for individualized diagnosis and treatment, and need for physician review; Future directions: Optimize data coverage, integrate patient individual characteristics, and comply with regulatory requirements.