Retrieval-Augmented Generation (RAG) systems rely heavily on retrieval quality to ground large language model outputs. In this work, we conduct a controlled ablation study comparing semantic retrieval, hybrid retrieval, and hybrid retrieval with Maximum Marginal Relevance (MMR) re-ranking on a stock-news question answering task. Using a fixed dataset snapshot and Gemini for both generation and automatic evaluation, we analyze retrieval relevance, grounding, and coverage. Our results show that hybrid retrieval improves coverage compared to semantic retrieval, while MMR further increases contextual diversity without significantly improving grounding. These findings highlight that retrieval diversity and grounding are distinct concerns in RAG system design.