FinAgent-RAG: Agentic Framework for Financial Document QA

A recent research article presents FinAgent-RAG, a framework for retrieval-augmented generation aimed at answering questions related to financial documents. This domain of financial QA necessitates intricate multi-step numerical reasoning across diverse evidence types, including structured tables, textual narratives, and footnotes found within corporate filings. Traditional RAG methods utilize a single-pass retrieve-then-generate strategy, which often falters in handling complex reasoning sequences. In contrast, FinAgent-RAG implements iterative retrieval-reasoning cycles with self-verification to enhance accuracy in financial numerical reasoning. Key innovations of the framework comprise a Contrastive Financial Retriever that employs hard negative mining, a Program-of-Thought module for generating executable Python code, and a self-verification component. The study can be accessed on arXiv with the identifier 2605.05409.