Intelligent Chinese patent medicine (CPM) recommendation framework: Integrating large language models, retrieval-augmented generation, and the largest CPM dataset

Chinese patent medicines (CPMs), a vital component of healthcare systems in China and worldwide, has been increasingly utilized in clinical practice. However, approximately 70 % of CPMs are prescribed by Western medicine physicians who lack expertise in traditional Chinese medicine syndrome differentiation and treatment. Here, in this study, we constructed RAG-CPMF, an intelligent CPM recommendation framework integrating large language models (LLMs), retrieval-augmented generation (RAG), and the largest CPM dataset. Specifically, we first applied the Multi-LLMs Validation method, significantly reducing the manual proofreading workload involved in structured data extraction. Furthermore, based on this approach, we successfully established the largest CPM dataset (https://gitee.com/tcmdoc/cpm) with continuous on-line updates. Afterwards, we combined the CPM dataset and the RAG architecture, along with effective integration of generative capacity from LLMs and external data retrieval; all of which contributed to the RAG-CPMF. Finally, the accuracy of RAG-CPMF was evaluated against clinical guidelines, demonstrating that this framework significantly improved CPM recommendation accuracy compared with general-purpose LLMs.