近年來，大型語言模型在自然語言處理領域取得顯著進展。然而，將其應用於高度結構化之產品資料表時，常因模型缺乏精確之邏輯推演與數值比對能力，導致嚴重的數值幻覺，難以滿足實務運營對資料真實性之嚴格要求。本研究之主要目的，旨在探討並建構一套高可靠度之結構化資料檢索系統，以克服現有生成式語言模型之技術侷限。 本研究提出一套「基於人工智能代理之兩階段結構化資料檢索框架」。系統第一階段利用本地端部署之程式碼生成模型，結合智能代理之工具調用能力，將自然語言轉譯為程式碼進行實體層面之過濾；並導入領域綱要映射與狀態邏輯。同時於代理框架中實作基於執行期錯誤日誌之自我修正機制。第二階段則將過濾後之數據轉化為自然語言總結，確保輸出結果具備絕對之真實性。 實驗結果顯示，本研究所提框架之整體執行準確率高達87.3%，大幅超越純文本語義檢索(24.0%)與直接生成式問答(40.6%)等基線系統。消融實驗進一步證實，自我修正機制能有效將語法錯誤率由31.4%收斂至4.6%。本研究成功驗證該人工智能代理框架能消弭模型幻覺，為結構化資料檢索提供具備高度穩健性與實務決策價值之解決方案。;In recent years， large language models (LLMs) have made significant advancements in the field of natural language processing. However， when applied to highly structured product data tables， these models often suffer from severe "numerical hallucinations" due to a lack of precise logical reasoning and numerical comparison capabilities， making it difficult to meet the strict data authenticity requirements of practical operations. The primary purpose of this study is to explore and construct a highly reliable structured data retrieval system to overcome the technical limitations of existing generative language models. This study proposes a "Two-Stage Artificial Intelligence Agent Structured Data Retrieval Framework." In the first stage， the system utilizes a locally deployed code-generation model， combined with the tool-calling capabilities of an AI agent， to translate natural language into code for physical-level filtering. Simultaneously， it introduces domain schema mapping and status-based fail-safe logic. Additionally， a self-correction mechanism based on runtime error logs is implemented within the agent framework. The second stage synthesizes the filtered data into a natural language summary， ensuring the absolute authenticity of the output results. Experimental results demonstrate that the overall execution accuracy of the proposed framework reaches 87.3%， significantly outperforming baseline systems such as pure semantic search (24.0%) and direct generative question answering (40.6%). Ablation studies further confirm that the self-correction mechanism effectively reduces the syntax error rate from 31.4% to 4.6%. This study successfully verifies that the artificial intelligence agent framework can eliminate model hallucinations， providing a highly robust solution with practical decision-making value for structured data retrieval.
