Cedric Thomas Y. Rafanan, Rhenz Gerard G. Largo, Glenn B. Pepito
Retrieval-augmented generation (RAG) enhances large language models (LLMs) by grounding their responses in external knowledge, which is crucial for domains such as finance where factual accuracy and interpretability are vital. However, conventional RAG pipelines often treat heterogeneous financial data comprising unstructured narrative text and structured tabular statements as a single modality, resulting in information loss and reduced retrieval precision. This study presents a hybrid RAG architecture that adapts retrieval methods to the data type. An adaptive retrieval router classifies queries and directs them to either a dense retriever (Qwen 0.6B) or a sparse retriever (BM25) with results integrated through a reranker for unstructured data, while a graph-based retriever built on Neo4j handles structured, multi-hop, and temporal reasoning. The system was evaluated on 17,000 questions derived from 200 U.S Securities and Exchange Commission (SEC) filings using multiple metrics. Results show that integrating a graph-based retriever improved context recall by 63%, answer correctness by 31%, and overall performance by 12% compared to flattened text retrieval. By addressing modality misalignment in financial RAG, the proposed system enhances factual accuracy, reasoning depth, and interpretability, establishing a scalable foundation for future research in graphbased financial document understanding. © 2026 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
