Analisis Perbandingan Metode Chunking Dalam Chatbot Berbasis Retrieval-Augmented Generation Rekomendasi Terapi Nutrisi Medis Pasien

Abstract

Artificial intelligence chatbots are increasingly used in healthcare services, yet the accuracy of their responses depends greatly on the effectiveness of the retrieval process in Retrieval-Augmented Generation (RAG) systems. This study compares three chunking methods: recursive chunking, semantic chunking, and Double Pass Merging Chunking, within a RAG-based chatbot designed to deliver medical nutrition therapy recommendations. The evaluation uses a dataset of medical nutrition reports and applies BLEU, ROUGE, and RAGAS metrics such as context precision, context recall, faithfulness, and answer relevancy. The analysis results show that Double Pass Merging Chunking delivers the best performance across 7 out of 8 evaluation metrics, achieving the highest scores in Context Precision (0.5514), Context Recall (0.2721), Faithfulness (0.2381), BLEU (0.9129), ROUGE-1 (0.1117), ROUGE-2 (0.0174), and ROUGE-L (0.0680). Semantic Chunking showed significant improvement in Context Recall (0.2702) with a 48% difference compared to Recursive Chunking. These findings highlight the substantial impact of chunking strategy selection on the accuracy of medical nutrition recommendations produced by RAG-based chatbots.