隨著針灸臨床應用與相關研究持續發展，針灸文獻數量日益增加。然而，穴位與疾病症狀之間的治療關聯往往分散於不同文獻中，當臨床工作者與研究者在查詢與整 合文獻時，往往需要耗費大量時間。大型語言模型 (Large Language Model， LLM) 雖具備良好的語意理解與文字生成能力，但若直接應用於醫療問答任務，仍可能受到知識更新、幻覺現象與證據來源不明等限制。為改善上述問題，本研究開發一套基於檢索增強生成技術 (Retrieval-Augmented Generation， RAG) 的針灸治療問答框架，藉由引入外部文件證據，降低模型生成缺乏依據或不準確內容的風險。 在系統架構上，本研究以 CNKI 與 VIP 資料庫作為主要資料來源，收集 42，389 筆同時包含穴位與疾病/症狀資訊之針灸摘要，作為系統之外部知識來源。接著，分別建置關鍵詞檢索和語意檢索，並採用加權倒數排名融合方法整合兩種檢索結果，以提升相關摘要之召回效果。系統評估方面，則以內部測試集與外部測試集進行分析，分別評估系統之檢索成效、生成品質與選擇題作答正確率。 研究結果顯示，在內部測試集中，檢索階段能有效找回與問題相關之摘要，生成階段則能依據檢索的證據進行回答，降低幻覺風險；在外部測試集中，引入 RAG 後作答表現皆有所提升，說明本研究建置之針灸治療相關文獻可作為檢索增強架構中的外部知識來源，有助於提升大型語言模型於針灸問答任務中的回答品質與正確率。;As the clinical application and related research of acupuncture continue to evolve， the volume of acupuncture literature is growing rapidly. However， the therapeutic relationships between acupoints and diseases or symptoms are often scattered across different studies， requiring clinicians and researchers to spend a significant amount of time searching for and synthesizing this information. Although Large Language Models (LLMs) possess strong semantic understanding and text generation capabilities， directly applying them to medical question-answering tasks may still be limited by issues such as outdated knowledge， hallucination， and unclear evidence sources. To address these issues， this study develops an acupuncture therapy question-answering framework based on Retrieval-Augmented Generation (RAG)， which incorporates external literature evidence to reduce the risk of generating unsupported or inaccurate content. In terms of system architecture， this study utilized the CNKI and VIP databases as primary data sources， collecting 42，389 acupuncture literature entries that contained information on both acupoints and diseases/symptoms to serve as the system’s external knowledge source. Subsequently， keyword retrieval and semantic retrieval were implemented separately， and a weighted Reciprocal Rank Fusion method was adopted to integrate the results of the two retrieval approaches and improve the recall of relevant literature. For system evaluation， analyses were conducted using internal and external test sets to assess the system’s retrieval performance， generation quality， and accuracy in answering single-choice questions. The results showed that， in the internal test set， the retrieval phase was able to effectively retrieve literature relevant to the query， and the generated responses were able to answer based on the retrieved evidence， thereby reducing the risk of hallucinations. In the external test set， the answering performance improved after introducing RAG， indicating that the acupuncture related literature constructed in this study serves as an external knowledge source within the RAG framework， and helps improve the answer quality and accuracy of large language models in acupuncture question-answering tasks.
