A HALLUCINATION-RESISTANT AGENTIC RAG SYSTEM WITH MULTI-STAGE RETRIEVAL AND RESPONSE VALIDATION

Abstract

—Large Language Models (LLMs) have shown great abilities in understanding and generating natural language. However, they can create hallucinations by providing factually incorrect or unfounded responses. To tackle this issue, this work suggests a hallucinationresistant Agentic Retrieval-Augmented Generation (RAG) system that combines multi-stage retrieval, smart re-ranking, and response validation. The system uses an agentic structure, where independent agents plan, retrieve, and refine information by utilizing external knowledge sources. A multi-stage retrieval pipeline brings together dense vector search, keyword-based retrieval, and query reformulation to enhance recall and coverage. The retrieved documents go through a re ranking module that prioritizes contextually relevant and high confidence evidence, which reduces noise and irrelevant information. To improve reliability, we add a response validation layer that uses techniques like crossverification, fact-checking against retrieved evidence, and confidence scoring. This ensures that the generated outputs are based on verifiable sources and lowers the risk of hallucinations. Additionally, the system includes feedback loops and iterative reasoning, allowing agents to update outputs in real-time. Experimental results show that this framework greatly boosts answer accuracy, contextual relevance, and reliability compared to traditional RAG systems. This research presents a scalable and modular setup for implementing trustworthy AI systems in important fields such as healthcare, finance, and legal applications.