The study presents a new TOSRR (tree-organized self-reflexive retrieval) framework that improves the answers of large language models (LLMs) to traditional Chinese medicine (TCM) questions. Traditional methods cannot capture the hierarchical nature of TCM knowledge, so TOSRR uses subject-predicate-object-text (SPO-T) units in a tree architecture and an iterative self-reflection mechanism for dynamic knowledge acquisition and verification. Performance was tested on questions from the TCM Medical Licensing Examination (MLE) and Classics Course Exam (CCE). When integrated with GPT-4, TOSRR increased absolute accuracy in MLE by 19.85% and recall accuracy in CCE from 27% to 38%. Expert evaluation showed an improvement of 18.64 points in safety, consistency, explainability, conformity and coherence. RAGA metrics confirmed better knowledge utilization, search accuracy, and noise robustness compared to standard RAG methods. The framework has potential for application in TCM teaching.