<![CDATA[The development of artificial intelligence-based chatbots has enabled the automated delivery of information. However, chatbots still face challenges in generating relevant answers to user queries. In addition, Retrieval-Augmented Generation (RAG) systems that rely on a single retrieval method have limitations in identifying the most relevant documents, which may affect the quality of the generated responses. Therefore, this study aims to implement a Hybrid Retrieval-Augmented Generation (Hybrid RAG) approach by combining BM25, Bi-Encoder, and Cross-Encoder models. The dataset used consists of 200 Bali tourism documents collected from web scraping, the Bali Central Bureau of Statistics (BPS), and Kaggle. The system was developed using BM25 as the initial retrieval stage, Bi-Encoder for semantic retrieval, Cross-Encoder for reranking, and Gemini as the answer generation model. The evaluation was conducted using Recall@k, Precision@k, and nDCG@k metrics to measure retrieval performance, as well as response time testing to evaluate system performance. The results show that the Hybrid RAG approach, utilizing BM25, Bi-Encoder, and Cross-Encoder, improves document retrieval quality through keyword-based matching, semantic similarity analysis, and reranking processes. Based on the findings, the Hybrid RAG approach was successfully implemented in a Bali tourism information chatbot and was able to provide relevant context to support the answer generation process.]]>
Copyrights © 2026
