Large language models face significant challenges in specialized domains where hallucinations are unacceptable under stringent accuracy requirements. This paper presents Retrieval-Augmented Generation Fusion (RAG-Fusion), a novel RAG system employing reciprocal rank fusion with multi-query generation to optimize knowledge retrieval and reduce hallucinations in financial customer service. Unlike traditional single-query RAG systems, our approach implements multi-query generation and Reciprocal Rank Fusion (RRF) based re-ranking mechanisms, integrating composite query construction and semantic vector weighting to improve recall of complex financial terminology. Through collaboration with China UnionPay Data (CUPD), we construct a comprehensive professional domain knowledge database addressing the shortage of specialized financial corpus. Experimental evaluation against ChatGPT-4 and experienced industry specialists demonstrates that our method achieves 86% accuracy compared to 45% of the baseline models, performs comparably to decade-experienced specialists while requiring only 22% of their response time, and excels at understanding deep semantics and open-ended questions in financial domain.