本研究旨在探討智慧製造機台資料查詢情境中，基於本地化小型語言模型（Small Language Models， SLMs）之 ReAct-based Agentic RAG 多來源資料檢（Multi-resource Data Retrieval）架構，對回答品質與系統效能之影響。研究以 Llama 3.1-8B 作為代表性本地化小型語言模型，首先使用 AI4I 2020 資料集建立單一資料來源知識庫，再進一步整合 AI4I 2020、AI4I_PMDI 與 Predictive Maintenance v3，建構多來源機台知識庫，以模擬智慧製造中異質且部分資訊不完整之機台資料整合與查詢情境。 本研究設計簡單任務與複雜任務兩類測試問題，並採用大型語言模型評分（LLM-as-aJudge）機制，從回答正確性、推理品質、檢索依據與輸出格式四個面向進行評估。為提升系統處理複雜查詢的能力，本研究導入 ReAct 框架，並透過推理、行動與觀察的迭代循環，使系統能自主完成多步驟資料檢索、數值計算、證據評估與機台比較等任務。 實驗結果顯示，在簡單任務中，傳統檢索增強生成（Retrieval-Augmented Generation， RAG）與 ReAct-based Agentic RAG 的表現相近；然而，在多來源知識庫的複雜任務中，ReAct-based Agentic RAG 將平均分數由 71.70 提升至 93.20，展現較佳的推理能力與回答品質。整體而言，ReAct-based Agentic RAG 結合多來源資料檢索架構，能有效提升本地化小型語言模型於智慧製造情境下處理多來源資料檢索與複雜機台查詢的能力。雖然此方法增加了 Token 使用量、工具呼叫次數及回應時間，但能大幅提升整體回答品質。;This study investigates the effects of integrating ReAct-based Agentic Retrieval-Augmented Generation (RAG) with multi-resource data retrieval on response quality and system performance for machine data retrieval in smart manufacturing using a localized Small Language Model (SLM). Llama 3.1-8B was adopted as the representative localized SLM. A single-resource knowledge base was first constructed using the AI4I 2020 dataset and was subsequently extended into a multi-resource knowledge base by integrating AI4I 2020， AI4I_PMDI， and Predictive Maintenance v3. This configuration was designed to simulate practical industrial scenarios involving heterogeneous and partially incomplete machine data. Two categories of benchmark tasks， namely Simple Tasks and Complex Tasks， were designed to evaluate system performance. The generated responses were assessed using an LLM-as-a-Judge framework based on answer correctness， reasoning quality， evidence grounding， and output format. To enhance the system’s ability to handle complex queries， the proposed approach incorporates the ReAct framework through an iterative reasoning–action–observation cycle， enabling multi-step data retrieval， numerical computation， evidence evaluation， and machine comparison. Experimental results show that Conventional RAG and ReAct-based Agentic RAG achieve comparable performance on Simple Tasks. However， under the multi-resource knowledge-base setting， ReAct-based Agentic RAG improves the average score on Complex Tasks from 71.70 to 93.20， demonstrating stronger reasoning capability and higher response quality for multi-resource data retrieval and complex machine-related queries. Although the proposed approach requires more tokens， tool invocations， and response time， it substantially improves overall response quality in smart manufacturing applications.
