工程時程網圖為進度規劃、控制與工期爭議判斷之重要依據，然而現行審查多仰賴人工經驗或少部分的軟體工具，且既有自動化工具亦多集中於可量化與規則式項目，對於施工邏輯、語意理解、契約一致性與工程合理性等非量化問題仍有不足。本研究旨在探討大型語言模型（Large Language Model， LLM）應用於建築工程時程網圖審查之可行性與限制。 本研究透過文獻回顧整理國內外時程網圖審查規範、自動化審查作法及工具與LLM專業應用挑戰，並將審查項目歸納為格式、完整性、正確性與合理性四類。研究以MS Project匯出之Excel資料作為模型輸入，測試LLM於時程網圖元素辨識與時程邏輯運算之基本能力，並以ChatGPT自訂GPT建構結合提示詞工程與檢索增強生成（RAG）之輔助審查系統，比較零樣本、少樣本與少樣本結合RAG之表現。 研究結果顯示，結合RAG後，模型能引用外部規範與專案資料，提升專業判斷依據與審查完整性。不過在處理排程邏輯計算相關問題時LLM尚不足以取代排程軟體與人工審查所累積的經驗，但具備輔助作業語意判讀、審查流程標準化與跨文件資訊整合之應用潛力。;Construction schedule networks are essential for schedule planning， control， and the assessment of time-related disputes. However， current schedule review practices still rely heavily on human experience or limited software， and existing automated tools primarily address quantifiable， and rule-based issues. Their ability to examine non-quantitative issues， such as construction logic， semantics understanding， contractual consistency， and engineering reasonableness， therefore remains limited capabilities. This study examines the feasibility and limitations of using Large Language Models (LLMs) for construction schedule network review. Through a literature review， this study examines schedule network review standards， automated review approaches， and challenges in the professional application of LLMs. The review items of the schedule network are categorized into four dimensions: format， completeness， correctness， and reasonableness. Using Excel data exported from MS Project as model input， this study evaluates the basic capabilities of LLMs in schedule network element recognition and schedule logic calculation. It further develops a review support system using ChatGPT Custom GPT， integrating prompt engineering and Retrieval-Augmented Generation (RAG)， and compares the performance of zero-shot， few-shot， and few-shot with RAG approaches. The research results indicate that， after integrating RAG， the model can reference external standards and project-specific information， thereby strengthening the basis for professional judgment and the completeness of the review. Nevertheless， when dealing with the issues related to scheduling logic calculations， LLMs are still insufficient to replace professional scheduling software or the experience accumulated through human review. Despite this limitation， they demonstrate considerable potential for assisting in activity semantic interpretation， standardizing the review process， and integrating information across multiple documents.
