Integrating Retrieval-Augmented Generation and Large Language Models for Financial Question Answering
Journal
International Conference on Intelligent Informatics and BioMedical Sciences, ICIIBMS
Start Page
9
End Page
11
ISBN (of the container)
979-833150166-2
Date Issued
2026-01-02
Author(s)
Chen, Yu-Jen
Chen, Ping-Han
Tung, Tzu-Chia
Chou, Yung-Chien
Chu, Yu-Chin
Du, Sian-Wun
Wang, Wei-Chien
Yang, Chung-Ming
Abstract
Customer service systems in the financial industry require accurate and effi-cient question-answering solutions. Traditional methods, such as rule-based chatbots, struggle with complex queries, while Large Language Models (LLMs) face challenges like hallucination and outdated knowledge. This study explores the effectiveness of Retrieval-Augmented Generation (RAG) in answering financial questions using various retrieval methods, including BM25, embedding models, and reranker models. Experimental results show that BM25 is the fastest but less accurate, while the Reranker Model achieves the highest accuracy (0.8933) at a high computational cost. The best balance is found by combining BM25, a Reranker Model, and Recursive Token Chunker, improving accuracy (0.92) while reducing execution time (2427 seconds). This approach enhances AI-driven financial services by providing reliable and up-to-date responses.
Event(s)
10th International Conference on Intelligent Informatics and BioMedical Sciences, ICIIBMS2025
Subjects
Financial Question Answering
Large Language Models (LLMs)
Retrieval-Augmented Generation (RAG)
Publisher
IEEE
Type
conference paper