A new preprint introduces VisR-RAG-Fin, a benchmark designed for evaluating multimodal retrieval-augmented generation (RAG) systems on visually rich financial documents. The work responds to the growing use of automated systems to process enterprise documents such as financial reports, which are typically created for human readers.
The abstract notes that these documents exhibit complex visual structure, combining dense text with tables, charts, and other visual elements. This makes them a challenging testbed for RAG systems, which must go beyond plain text and account for layout and visual information to retrieve and generate accurate answers.
By focusing on this specific document type, VisR-RAG-Fin highlights a gap in existing RAG evaluation and provides a resource for measuring progress on multimodal understanding in the financial domain. The paper is available on arXiv under the identifier 2610.24822.