This study presents the development of a conversational culinary recommendation system in Indonesian using a Large Language Model (LLM). The system aims to generate personalized restaurant recommendations through natural language interaction while addressing the limited availability of Indonesian language conversational recommender systems. Data were collected from the online platform and enhanced with synthetic conversational data to train the model’s dialogue understanding and reasoning capabilities. The proposed architecture consists of four main components: Input/Output, Recommender Agent, Tools, and Memory, which collaborate to process user input, retrieve relevant restaurant information, and generate contextually appropriate responses. The Mistral 7B model was fine-tuned using supervised training and evaluated against Komodo 7B as a baseline model. Experimental results show that Mistral 7B achieved higher performance across all evaluation metrics, with a Hit@5 of 0.99, Average Turns@5 of 2.48, Recall@5 of 0.22 percent, and NDCG@5 of 0.20, surpassing Komodo 7B in both single turn and multi turn dialogue scenarios. These findings demonstrate that Mistral 7B provides greater stability, stronger contextual understanding, and higher efficiency in generating relevant and accurate restaurant recommendations in Indonesian.
Mauludin et al. (Thu,) studied this question.