Yidu Tech releases TCM evidence agent Xiaohuangqi: answers traceable sentence by sentence, 8,000+ terms translated

TCM clinical questioning gains an agent built around evidence tracing: a doctor asks a professional question in natural language, the system identifies symptoms, syndromes, treatment methods and formulas, then retrieves evidence from classics, clinical guidelines, expert consensus and modern medical research to give supporting analysis. Previously doctors had to integrate knowledge across classics, guidelines, consensus and literature themselves, while TCM knowledge is complex, its terminology distinctive and its evidence sources diverse, leaving traditional experience and modern evidence poorly connected. The agent was released by Dr. Li Linfeng, Yidu Tech's vice president of technology innovation and AI architect, at the results-release session of the Beijing International Traditional Medicine Exhibition during the 2026 World Conference on Traditional Medicine, and was demonstrated at the conference booth.

The system emphasizes evidence-grounded answers and traces each sentence to its source, so doctors can check classic texts, guidelines and medical literature to verify conclusions. The reliability of generated answers is a precondition for clinical adoption, and sentence-level tracing ties conclusions to verifiable original text in order to reduce the risk of AI hallucination. Its knowledge base includes more than 10,000 TCM knowledge entries from People's Medical Publishing House and more than 8,000 international TCM terms set by People's Medical Publishing House, the World Health Organization and the World Federation of Chinese Medicine Societies, supporting Chinese-English translation of syndromes, treatments and formulas; the source also says the entries include more than 11,000 ancient and modern case records, which is inconsistent with the 10,000-plus total, and both figures are kept as stated.

The evidence capability extends into the clinical workflow: the TCM clinical assistant Copilot is embedded lightly in hospital physician workstations and, using the current patient record and care context, calls the evidence engine and specialty task agents. A standalone evidence tool addresses only knowledge retrieval, whereas Copilot attaches the same capability to concrete clinical tasks, extending from knowledge access to record generation, diagnostic assistance, condition assessment and risk management. In TCM outpatient record generation, Copilot helps organize symptoms, tongue and pulse findings and related information into structured records; Yidu ZhiXun and the Yidu clinical assistant are reported in use at more than 40 well-known hospitals nationwide, providing intelligent assistance and evidence support for clinical care across specialties.

The report covers the release of Xiaohuangqi and the TCM clinical assistant Copilot at the 2026 World Conference on Traditional Medicine, the booth demonstration, and the stated use of Yidu ZhiXun and the Yidu clinical assistant at more than 40 well-known hospitals nationwide; it is written for researchers and engineers following clinical AI and TCM informatization.

Sentence-level tracing is presented as a means of reducing hallucination risk, and no quantified hallucination or accuracy rate is given; the use at more than 40 hospitals is also not broken down by specialty or task. The source calls the knowledge base more than 10,000 TCM knowledge entries and also says it includes more than 11,000 ancient and modern case records, a subset larger than the total, so the counting basis for the two figures remains unclear.