An Iterative Self-Correcting Agentic RAG System

Abstract

Traditional Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) with external knowledge but are limited by static, single-pass retrieval, making them less effective for complex, multi-step reasoning. This paper introduces a context-driven agentic RAG architecture that uses autonomous agents to dynamically manage retrieval, reasoning, and response generation. The system includes three specialized agents: a CorrectiveAgent for iterative, self correcting retrieval; a Pre-Act agent for query decomposition and structured reasoning; and a WorkflowAgent for task execution. An LLM-based router selects the most appropriate agent based on query context. The framework integrates web search tools and vector-based retrieval for up-to-date knowledge access. Evaluated using Gemini 2.5 Flash Lite, the routing mechanism achieved 97.96% accuracy on 49 queries. Further testing on 120 cases using DeepEval showed significant improvements over a standard RAG baseline, including higher Faithfulness (0.95 vs. 0.79), Answer Relevancy (0.42 vs. 0.24), and Contextual Relevancy (0.28 vs. 0.19). These results demonstrate that agentic RAG systems outperform traditional approaches, with potential for further gains using more advanced models or higher reasoning budgets.