Yaobao: An evidence-based Chinese patent medicine assistant based on large language models and multi-source datasets
Year: 2026
Type: article
Abstract: Background: Chinese patent medicine (CPM) has gained increasing attention for its clinical value, but knowledge resources remain fragmented and practical digital aids are limited. Objective: To develop Yaobao, a CPM knowledge assistant whose contributions lie in domain-specific data integration and application engineering using established retrieval-augmented generation (RAG) methods. Methods: Six types of CPM resources were integrated and transformed into 95,197 question-answer pairs for a dual-level knowledge base linking semantic retrieval with source metadata. Yaobao incorporates follow-up questioning, verifiable citations, multilingual support, multimodal input, and external tool invocation. Yaobao uses DeepSeek-V3 as its generator; comparison with the standalone DeepSeek-V3 model helps assess the contribution of CPM-specific retrieval. It was compared with five general-purpose and two traditional Chinese medicine (TCM)-specific large language models on a 1,000-item single-turn benchmark, using automatic text-overlap metrics and Gemini-2.5-Pro to assess how well responses covered the reference answers. Results: Yaobao achieved the highest ROUGE-2, ROUGE-L, and METEOR scores across all three question categories and the highest ROUGE-1 scores for efficacy and prescriptions. For ROUGE-1 on precautions questions, Qwen3-235B scored 0.6771 and Yaobao scored 0.6759, while Biancang-7B achieved the highest BLEU-1 scores in all categories. Yaobao received the highest overall mean score from Gemini-2.5-Pro (3.57). Conclusions: Yaobao integrates CPM knowledge and interactive consultation features through established RAG components. On the 1,000-item benchmark, it performed strongly on questions about CPM knowledge represented in its retrieval corpus, providing a basis for further evaluation of its consultation functions.
Source: AI Evidence-Based Medicine
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7226021263
