This paper compares Unsupervised and Supervised Retrieval-Augmented Generation (RAG) models and their differences in retrieval mechanisms, model architectures, and output variability. RAG models, which integrate retrieval and generation processes, overcome the shortcomings of conventional language models by improving factual accuracy and contextual relevance through external information retrieval. The research goes into the changing landscape of RAG, following the development from initial retrieval-based models to joint training of retrieval and generation modules. This work discusses unsupervised RAG methods, which use unlabeled data for dynamic adaptation, and supervised methods, which maximize precision using labeled datasets and task-specific supervision. This work contributes to analyzing the merits and limitations of the learning approaches, considering scalability, flexibility, and reliability. It expounds on their uses in knowledge-intensive tasks like open-domain question answering and sentiment analysis. The work further outlines significant techniques used in early works, illustrating improvements and drawbacks to shed light on the future of RAG models in NLP. It is concluded that Supervised RAG methods show higher accuracy while unsupervised approaches offer scalability and adaptability.