Agentic Retrieval-Augmented Generation (RAG) Framework with Quadruple-Based Reasoning and Reinforcement Learning (RL) Optimization

Abstract

Retrieval-Augmented Generation (RAG) has emerged as an effective technique to reduce hallucinations in large language models (LLMs) but they follow a static retrieve-then-generate pipeline. This process is insufficient for complex financial question answering system that require multi-step reasoning, numerical precision, and factual verification. Therefore, in this research, we proposed an RL-Driven Agentic Multi-HyDE RAG framework designed to improve factual correctness and informativeness through structured reasoning and reinforcement learning optimization. The proposed methodology comprises of six major components: query diversification, hypothetical answer generation (HyDE), dense embedding-based retrieval, quadruple-based atomic knowledge representation, reinforcement learning-based evaluation, and tool-augmented refinement. Experimental evaluation on financial queries using Sentence Transformers, FAISS, and Mistral-7B-Instruct demonstrates that the framework achieves high factual alignment (faithfulness score = 1.0) while maintaining informativeness, without unnecessary calling of the external tools. The results indicate that integrating agentic reasoning, structured knowledge extraction, and reinforcement learning significantly overcomes hallucinations and improves reliability. The proposed architecture provides a scalable and robust solution for high-stakes financial question answering systems.