Retrieval-Augmented Generation (RAG) enhances the capabilities of Large Language Models (LLMs) by integrating external knowledge retrieval into the generation pipeline, enabling responses that are grounded, adaptive, and up to date. While RAG can improve factual accuracy compared to models relying solely on pre-trained data, its effectiveness in practice depends strongly on the quality, relevance, and interpretability of retrieved context, and does not eliminate hallucinations entirely. Recent architectures such as Fusion-in-Decoder, Atlas, and ColBERT-RAG demonstrate measurable gains in retrieval precision, scalability, and cross-domain generalization. However, persistent challenges remain, including retrieval noise that can override model reasoning, hallucinations that persist even with high-quality evidence, latency constraints that hinder real-time deployment, and fragile domain adaptation. Moreover, although partial metrics and task-specific benchmarks exist, the absence of a unified evaluation framework for retrieval–generation grounding and robustness complicates fair comparison and reproducible progress. Rather than offering an exhaustive survey, this review provides a focused analytical perspective on retrieval design and architectural evolution in RAG systems. It consolidates representative architectures while critically examining structural limitations related to retrieval–generation coupling as a design choice, context over-reliance, and privacy-preserving computation. Building on these insights, the paper outlines future research directions, including structured knowledge integration via GraphRAG, modular agent-based orchestration in Agentic RAG, improved retrieval filtering, and unified evaluation methodologies. As RAG architectures continue to evolve rapidly in a pre-standardization phase, a more analytically grounded understanding of their design trade-offs is essential for advancing trustworthy and domain-adaptive language systems