<p>Recent advancements in large language models (LLMs) have transformed conversational AI in healthcare, offering promising solutions for question answering (QA) tasks. However, critical challenges such as hallucinated content, dependence on outdated or incomplete knowledge, and the high computational cost of fine-tuning continue to hinder their effective deployment, particularly in resource-constrained settings. Retrieval-augmented generation (RAG) addresses some of these limitations by integrating external knowledge into the generation process. To fully harness its potential, optimized retrieval strategies tailored to the healthcare domain are required. To address this gap, this paper systematically investigates the performance of RAG-based LLMs for healthcare QA and proposes an ensemble retrieval framework augmented with a re-ranking technique to enhance the reliability and factual consistency of generated responses. The framework has been tested on a range of medical QA datasets with iterative systematic reviews using standard performance metrics such as accuracy, precision, recall, and F1-score. Additionally, expert validation by healthcare professionals ensures clinical relevance and practical applicability. As a baseline, model such as BioMistral is evaluated on the same dataset to highlight the limitations of standalone generation and emphasize the impact of RAG in improving the response generation performance without extensive fine-tuning. This paper offers new insights into the capabilities of RAG-enhanced LLMs in generating trustworthy responses without the need for fine-tuning, underscoring their potential as scalable and cost-effective solutions for real-world, AI-driven healthcare question answering systems.</p>