Large Language Models (LLMs) provide transformative capabilities which impact healthcare as well as cybersecurity and natural language processing and other sectors of application. Large Language Models create hazardous conditions when they hallucinate because their fabricated outputs produce wrong and misleading information. The research presents an innovative Learning Loop Management structure which combines Red Teaming with Prompt Engineering and Retrieval-Augmented Generation (RAG) to reduce occurrences of hallucinations. The model achieves better reliability in processing multimodal information through its connection between adversarial testing alongside CNN-based context-aware prompting. Synthetic data experiments utilized multiple modes for testing while achieving test accuracy of 0.75 at 50 epochs and ROC-AUC of 0.93 combined with F1-score of 0.85 and a BLEU score of 0.94. Advanced preprocessing approaches with CNN-based architecture establish powerful solutions to remove noise and correct data imbalance which enhances the factual accuracy in LLMs.