A Pengembangan Chatbot Informasi Kesehatan Ibu dan Anak Jawa Barat Berbasis Hybrid RAG dan TextToSQL

Abstract

The high Maternal Mortality Rate and stunting prevalence in West Java hinder the achievement of SDGs. Information accessibility remains a key constraint as data is scattered in narrative formats within Health Profiles and MCH Books. This study develops a hybrid chatbot based on Retrieval-Augmented Generation (RAG) and Text-to-SQL using the Google Gemini Large Language Model. The system integrates hierarchical chunking, intelligent routing, conversational memory, and automatic data visualization to present factual and statistical information precisely. System evaluation demonstrates high performance: Text-to-SQL reached 93.33% execution accuracy and 85% Router accuracy. In the RAG module, testing on 60 questions yielded scores of 0.990 Faithfulness, 0.883 Context Recall, and 0.950 Answer Relevancy. The system achieved a perfect score of 1.000 for out-of-context handling, proving safety from hallucinations. User Acceptance Testing (UAT) recorded a 80% success rate, where the chart feature was rated significant in aiding data understanding. The study concludes that the hybrid approach effectively enhances health insight accessibility for the public and stakeholders, although complex table extraction requires further optimization.