The spread of fake news on social media is extremely quick and depletes the level of trust among the population. It is hard to detect fake news in the Bangla language due to ambiguity in sentence meaning, sarcasm, and the use of images with misleading text. The old approaches to machine learning and early deep learning are typically unable to cope with these multimodal difficulties. In this paper, a Large Language Model (LLM)-enhanced Retrieval-Augmented Generation (RAG) architecture is suggested to detect fake news. The system uses ChromaDB to retrieve the relevant information, the Serper API to verify news claims in real-time by using the web search service, and the EasyOCR to extract Bangla text in pictures. To enhance reliability and minimize false or hallucinated responses, two large language models, Gemini 2.5 Flash and Groq LLaMA-3.1, are combined and applied jointly and with a fusion-based decision process. The proposed framework is evaluated on the BanFakeNews dataset consisting of more than 57,000 Bangla news articles, and on a larger balanced assessment sample of 500 samples of 250 fake and 250 real cases of text and image modalities. Based on the experimental results, the proposed dual-model framework yields an 87% accuracy, 87.8% precision, 86% recall, and 86.9% F1 score, which is higher in comparison with the proposed Gemini 2.5 Flash model (84% accuracy) and Groq LLaMA-3.1 model (82.2% accuracy). The findings indicate that the retrieval-based evidence—LLM reasoning becomes an effective and reliable method in the detection of fake news in Bangla.