The phenomenon of “hallucination" remains the biggest barrier to the reliability of Retrieval-Augmented Generation (RAG) systems, especially when large language models (LLMs) synthesize information from multiple document sources. Traditional RAG pipelines often prioritize optimizing retrieval relevance but overlook controlling the factual consistency of the final answer. To tackle this problem in a more practical way, we introduce an enhanced Agentic RAG framework that comes with a built-in Fact-Verification Feedback Loop. In this setup, a dedicated agent checks each statement generated by the model against the retrieved evidence to make sure it actually holds up. Whenever the agent finds something inconsistent or lacking support, it automatically starts a correction cycle—refining the query and running the retrieval-generation process again until the answer reaches the required level of factual accuracy. The main objective of the proposed method is to directly reduce the hallucination rate and enhance the factual fidelity of RAG-based responses. This study contributes an agent-based workflow with strong self-regulation capabilities, promising to significantly improve the reliability of LLM applications in real-world scenarios.