Mafin 2.5 achieves 98.7% accuracy in FinanceBench by changing the retrieval query.

AIMPACT News, May 4th (UTC+8), Vectorless RAG improves retrieval accuracy by removing the vector database. The PageIndex style system changes the traditional retrieval logic, no longer asking “which text block is closest to the query,” but directly locating “where the answer is in the document.” This change allows the Mafin 2.5 system to achieve a 98.7% accuracy rate on the FinanceBench benchmark, and this achievement is not dependent on a better embedding model; it is the change in retrieval primitives that brings the performance boost. Traditional vector RAG systems have inherent limitations when handling complex nested queries (such as “Q3 2024 operating revenue excluding restructuring costs”) and precise numerical queries, which is the core reason why Vectorless RAG can achieve more accurate retrievals.