MIAT方法論為一套具備邏輯化、結構化的複雜系統設計流程，涵蓋以IDEF0拆解系統架構、使用Grafcet進行離散事件建模，最終透過高階合成產生程式碼，對於複雜系統設計具有明確且有效的指導作用。然而，大型語言模型（LLM）在應用MIAT方法論時易產生幻覺問題，且Vibe Coding風潮亦引發一系列程式碼生成品質疑慮。為此，本研究建置了以IDEF0階層概念設計的MIAT方法論知識圖譜（MIAT-KG），並導入Docling工具輔助建置過程。在知識理解能力測試中，導入知識圖譜後LLM於MIAT核心概念掌握的加權準確率由18%提升至93%；在系統設計輔助測試中，LLM能依循MIAT方法論的標準流程，逐步引導或啟發使用者完成系統架構拆解與離散事件建模，最終生成具備階層式、模組化結構的程式碼。且針對Vibe Coding與MIAT方法論生成之程式碼進行四種大型語言模型的評分比較，結果顯示MIAT方法論生成之程式碼在結構性與可維護性上均顯著優於Vibe Coding版本。MIAT方法論不僅確保程式的邏輯一致性與穩定性，也大幅提升後續系統擴充與維護的便利性。綜合實驗結果MIAT-KG能有效強化LLM於複雜系統設計中之輔助能力，提升程式生成之專業性與可控性，對於中大型系統開發具重要應用價值。;The MIAT methodology is a logical and structured design process for complex systems, encompassing system architecture decomposition using IDEF0, discrete event modeling through Grafcet, and final high-level synthesis to generate code, providing clear and effective guidance for complex system design. However, large language models (LLMs) often encounter hallucination issues when applying the MIAT methodology, and the Vibe Coding trend has also raised concerns regarding code generation quality. To address these challenges, this study constructed the MIAT Knowledge Graph (MIAT-KG) based on the hierarchical concepts of IDEF0 and utilized the Docling tool to assist in the construction process. In the knowledge comprehension evaluation, the introduction of the knowledge graph improved the LLM’s weighted accuracy in mastering MIAT core concepts from 18% to 93%. In the system design assistance tests, LLMs were able to follow the standard MIAT process, gradually guiding or inspiring users to complete system architecture decomposition and discrete event modeling, ultimately generating code with a hierarchical and modular structure. Furthermore, a comparative evaluation between Vibe Coding and MIAT methodology-generated code using four different LLMs showed that code generated following the MIAT methodology significantly outperformed that of Vibe Coding in terms of structure and maintainability. The MIAT methodology not only ensures logical consistency and system stability in code generation but also greatly enhances the ease of future system expansion and maintenance. Based on comprehensive experimental results, MIAT-KG is demonstrated to effectively strengthen LLM capabilities in assisting complex system design, enhancing the professionalism and controllability of code generation, and providing significant value for medium- to large-scale system development.
