<![CDATA[Large Language Models are increasingly adopted as planning components that translate natural language into structured representations for tool invocation, enabling executable interaction with backend systems through JSON based function calling. However, empirical studies focusing on Indonesian natural language remain limited. This paper presents a restaurant search chatbot case study that investigates JSON based function call planning from Indonesian user queries, with emphasis on the upstream planning task rather than conversational response generation. A synthetic dataset of 33,470 Indonesian restaurant search queries paired with ground truth JSON plans was constructed based on a predefined tool set and database schema. Supervised fine tuning with parameter efficient adaptation was applied to a pretrained language model. The fine tuned Mistral 7B model was evaluated using multiple metrics measuring JSON structural validity, tool sequence correctness, and parameter accuracy at different granularities. The results show strong performance, achieving a JSON structure validity rate of 0.97, tool sequence exact match accuracy of 0.92, column level accuracy of 0.97, and value level accuracy of 0.94. More stringent evaluation at the session level reveals remaining challenges in composing all parameters correctly within a single planning instance. Overall, the findings demonstrate that with carefully designed datasets and strict supervision, Large Language Models can reliably perform structured JSON based function call planning from Indonesian natural language, providing a practical foundation for extending this approach to other structured application domains where execution correctness is critical.]]>
Copyrights © 2026
