Traditional medical research is an important application scenario for natural language processing, and the differentiation of the Traditional Chinese Medicine (TCM) syndrome is an important challenge among them. We research the TCM syndrome differentiation process proposed in Evaluation 1 of the 10th China Conference on Health Information Processing (CHIP 2024) and propose a TCM syndrome differentiation method based on iterative retrieval-enhanced generation technology. A general large language model can achieve relatively high accuracy in a small-sample knowledge database. The method organically combines pathogenesis reasoning and syndrome reasoning, and the two work together to refine and improve the accuracy of syndrome differentiation. The experimental results on the test dataset show that the method proposed in this paper is effective, with a final score of 32.0534, ranking fifth in this task.