A knowledge-enhanced prescription generation model for TCM syndrome differentiation via multilevel memory retrieval

Abstract

This paper presents a prescription generation model for Traditional Chinese Medicine (TCM) based on multi-level memory retrieval and knowledge enhancement. The proposed approach addresses common limitations of large language models(LLMs) in TCM applications, such as outdated knowledge and hallucinations. The model integrates three key data sources: a structured clinical case database, a drug knowledge graph built from the Compendium of Materia Medica, and classical compatibility principles, including the “Eighteen Incompatibilities” and “Eighteen Antagonisms.” A hierarchical retrieval framework is constructed to leverage these resources effectively. Specifically, a multi-feature weighted BM25 algorithm retrieves similar cases from the database. Core drugs are extracted from the corresponding prescriptions, and additional retrieval identifies their main indications and compatibility constraints. This information—comprising retrieved cases, drug knowledge, and compatibility rules—is fed into an LLM using a carefully designed prompt to guide the generation of safe and clinically relevant prescriptions. Experiments conducted on the CCL2025 TCM benchmark demonstrate that the proposed model significantly improves the quality and reliability of prescription generation through multi-level retrieval and knowledge integration.