<p>With the rapid development of Large Language Models (LLMs), their application in Traditional Chinese Medicine (TCM) has become increasingly feasible. LLMs can integrate classical TCM texts, contemporary research findings, clinical case records, and expert knowledge to generate comprehensive and interpretable responses. However, LLMs are prone to “hallucinations,” producing plausible but incorrect diagnoses or prescriptions, which can pose serious risks in clinical decision making. To address this challenge, we propose TCMKG, an innovative knowledge-driven question answering framework that combines LLM-based reasoning with a meticulously constructed TCM knowledge graph. The core innovation of TCMKG is its two-stage framework: (1) An advanced prompting pipeline has been developed for the Xin’an Medical Case dataset, significantly enhancing the precision of entity extraction from unstructured TCM texts through expert-guided prompting strategies; (2) A large-scale domain-specific TCM knowledge graph containing 37,192 entities and 130,030 relations has been constructed based on a manually curated corpus. By integrating this knowledge graph with LLMs through Cypher-based querying, we have developed a hybrid approach that grounds the generation process in factual medical knowledge, thereby enabling controllable, interpretable, and trustworthy TCM question answering. The results show that TCMKG has achieved significant improvements in information extraction, with a precision of 0.89, recall of 0.88, and an F1 score of 0.88. In the knowledge-based question-answering task, TCMKG achieved an average RAGAS score of 0.954, while expert evaluations yielded an average score of 2.84, significantly outperforming baseline large language models.</p>