This paper presents a comparative evaluation of two approaches for developing a medical assistant chatbot: Retrieval-Augmented Generation (RAG) and fine-tuning. Both aim to improve generative modeling with domain-specific knowledge, but differ in performance and adaptability. Fine-tuning involves adapting a large language model (LLM) using a domain-specific dataset. It performs well in static environments but requires frequent updates to stay relevant. In contrast, RAG dynamically retrieves external documents during inference, enriching outputs with current knowledge without retraining the model. RAG is particularly well-suited for the medical domain, where information evolves rapidly. Its ability to incorporate up-to-date content enables more accurate and context-aware responses. Compared to fine-tuning auto-regressive transformers, RAG shows better performance in response quality, notably by reducing hallucinations, and offers superior scalability and lower maintenance costs. The paper concludes with future directions, including hybrid models combining RAG with fine-tuning systems, the integration of Agentic RAG architectures to enhance case-based reasoning, clinical decision-making, and real-time medical surveillance through iterative retrieval and reasoning mechanisms, and Self-supervised retriever adaptation to enhance performance and personalization without manual annotation.