Retrieval-Augmented Generation for Financial Question Answering over Long Documents
Amine Kobeissi (amine (point) kobeissi <at> umontreal (point) ca)
RALI, DIRO
Le 15 avril 2026 à 11 h 30
Salle 3195, Pavillon André-Aisenstadt — En présentiel, avec diffusion simultanée sur Zoom
Retrieval-augmented generation (RAG) is increasingly used for financial question answering over long regulatory filings, yet its reliability depends on retrieving the exact context needed to justify answers in high-stakes settings. Financial documents present unique challenges due to their length, structure, and presence of multiple data sources required in answering questions. However, existing benchmarks only partially capture these challenges. This presentation focuses on the retrieval component of the RAG pipeline, we evaluate a suite of retrieval strategies including dense, sparse, hybrid, and hierarchical methods with reranking and query reformulation. We identify where current bottlenecks exist by analyzing why these systems fail, specifically highlighting a frequent failure mode where the correct document is found but the specific context required to answer is missed. We tackle the problem by proposing a variation of a hierarchical approach. Based on empirical results from baseline methods and our proposed solution, we investigate three central questions. (1) What are primary bottlenecks in Financial RAG systems? (2) How do retrieval challenges and success rates differ across retrieval methods? (3) What characteristics make financial queries particularly challenging for retrieval? We conclude by outlining future research directions.
Pour recevoir les annonces hebdomadaires par courriel, visitez http://rali.iro.umontreal.ca/rali/?q=fr/node/1631