Building Indonesian Food Database with LLM

Session Track: App Dev

Session Time: 05:15 am - 05:30 am November 6

Session description

The rapid advancement of Large Language Models (LLMs) such as Gemini offers a transformative opportunity to convert unstructured or semi-structured data into meaningful graph representations. In this project, we present a practical application of LLMs to build a Nutrition Knowledge Graph from a publicly available Indonesian food dataset. The dataset, originally in CSV format, contains nutritional values such as calories, protein, fat, and carbohydrates for various local food items. We developed a pipeline that leverages Gemini to extract RDF-style triples from each data row via prompt engineering, transforming tabular entries into semantic relationships. These triples are then parsed and stored in a Neo4j graph database using Cypher queries, enabling rich querying and visualization of nutritional connections. The system architecture is built with Python (in Jupyter Notebook) and integrates Gemini AI. The resulting knowledge graph enables use cases such as dietary recommendation engines, semantic search for nutritional needs (e.g., “foods high in protein but low in fat”), and future extensions like health-aware or cuisine-specific knowledge graphs. This work demonstrates a scalable and low-code method for bootstrapping domain-specific knowledge graphs using LLMs, highlighting its potential for broader applications in health, agriculture, and education.