Agentic retrieval-augmented generation for financial QA: a SingleStore optimization approach

As the adoption of large language models (LLMs) accelerates, many small and medium enterprises lack the resources needed to train and maintain domain-specialized systems. Retrieval augmented generation (RAG) architectures provide a cost-effective alternative by leveraging external knowledge bases for enhanced performance. However, in high-stakes domains such as finance, ensuring accuracy and reliability remains a significant challenge. This paper introduces an Agentic RAG framework designed to improve financial question answering by integrating an adaptive, multi-step retrieval process within a Single Vector Store (SingleStore) setting. Utilizing a benchmark of financial questions (FinanceBench), our experiments show a jump in accuracy from the previously reported 50% to around 70% in single-store retrieval. The iterative retrieval and reasoning approach mitigates many common issues such as incomplete context and hallucinations while preserving the efficiency benefits of SingleStore. The results underscore the potential for more robust, cost-efficient financial chatbots tailored to the specialized needs of financial institutions.