隨著大型語言模型逐漸應用於語音助理與工具呼叫任務，語音助理已不再只是辨識使用者意圖，而是需要將自然語言指令轉換為可執行且安全的工具呼叫。然而，在邊緣車用環境中，系統受到運算資源、延遲與隱私需求限制，難以完全依賴雲端大型模型或針對每項任務重新訓練模型。因此，如何在不微調本地大型語言模型的情況下提升工具呼叫穩定性，成為重要研究議題。 本研究提出一套面向邊緣車用語音助理之免訓練代理式工具使用框架。此框架透過 ToolCard metadata 描述工具功能、參數規格、別名與安全規則，並結合 Tool Retriever、Planner、Function Caller、Argument Normalizer、Verifier 與 Replanning 等模組，將工具呼叫流程拆分為可控步驟，以提升工具選擇、參數生成、格式驗證與安全拒絕能力。 為評估本研究方法，本研究建立 Edge-IV-400 車用工具使用評估集，共 400 筆測試資料，涵蓋 easy、medium、hard 與 safety 四種難度。實驗結果顯示，完整框架 full_topk3 在整體正確率上達到 0.8600，且在明確單工具指令、口語化參數、安全拒絕與多工具規劃等任務中展現穩定性。消融實驗亦顯示，Argument Normalizer 與 Replanning 對整體表現具有明顯幫助。整體而言，本研究證明透過 metadata-driven 與 agentic verification loop 的設計，即使不重新訓練模型，仍可提升本地大型語言模型於車用工具呼叫任務中的可執行性與穩定性。;As large language models are increasingly applied to voice assistants and tool-use tasks， voice assistants are required not only to understand user intent but also to convert natural language commands into executable and safe tool calls. However， edge-oriented in-vehicle environments are often constrained by computation resources， latency， and privacy requirements， making it difficult to rely entirely on cloud-based large language models or to fine-tune models for every vehicle function. This thesis proposes a training-free agentic tool-use framework for edge-oriented in-vehicle voice assistants. The proposed framework uses ToolCard metadata to describe tool functions， parameter schemas， aliases， and safety rules. It further integrates Tool Retriever， Planner， Function Caller， Argument Normalizer， Verifier， and Replanning modules to decompose the tool-calling process into controllable steps. Through this design， the system aims to improve tool selection， argument generation， schema validation， and safety refusal without modifying the weights of the local language model. To evaluate the proposed method， this thesis constructs Edge-IV-400， an in-vehicle tool-use evaluation set containing 400 test cases across easy， medium， hard， and safety difficulty levels. Experimental results show that the complete full_topk3 framework achieves an overall accuracy of 0.8600. The system demonstrates stable performance on explicit single-tool commands， spoken argument expressions， safety refusal， and multi-tool planning tasks. Ablation results further indicate that Argument Normalizer and Replanning contribute significantly to the overall performance. These results show that a metadata-driven design with an agentic verification loop can improve the executability and stability of local large language models for in-vehicle tool-calling tasks without model fine-tuning.
