Large Language Models (LLMs) exhibit impressive fluency but remain prone to hallucinations-factual inaccuracies that hinder real-world deployment. We present AutoRAG-LoRA, a modular Retrieval-Augmented Generation (RAG) framework that mitigates hallucinations through lightweight LoRA-based adapters and KL-regularized fine-tuning. Our system combines structured prompt rewriting, hybrid sparse-dense retrieval, and hallucination detection using both classifiers and self-evaluation. When hallucination risk is detected, a feedback correction loop activates LoRA adapters to enforce factual alignment via contrastive KL loss. AutoRAG-LoRA operates efficiently under low-VRAM constraints and enables rapid adaptation without full model retraining. Empirical results on TruthfulQA, FEVER, HotpotQA, and an in-house adversarial set show that AutoRAGLoRA reduces hallucination rates by 46.6% over vanilla RAG.