Enhancing Large Language Models with External Knowledge Using Retrieval-Augmented Generation: A Critical Review of Retrieval Design, Context Compression, Safety and Evaluation, with a Conceptual Governance-Aware Architecture

Large language models (LLMs) depend on static parametric knowledge, leading to outdated answers, missing expertise, and hallucinations. Retrieval-Augmented Generation (RAG) addresses these issues by retrieving external evidence dur-ing inference, but the interactions between design choices are not well understood. This critical review explores configu-ration benchmarking, embedding quality, question-type-adaptive retrieval with context compression, comparison with fine-tuning, IAM-based permission filtering, and real-time, safety-critical graph-based RAG. A focused narrative design was used on a defined corpus, with foundational sources. The findings show that no single RAG configuration is best, as the optimal choice depends on the task and metrics. Retrieval quality strongly correlates with answer quality for ex-tractive tasks but weakly for synthesis tasks. Adaptive retrieval and compression can reduce token usage but increase latency. Access control, freshness, and safety can be enforced outside the LLM, but this is only possible under explicit trust assumptions. However, most studies rely on Wikipedia-based question answering, small samples, single runs, and inconsistent metrics. Several studies also contain internal inconsistencies, which are documented here. This review con-tributes five key elements: a survey matrix, a comparative analysis that separates different experimental settings, an eight-layer conceptual architecture justified by literature, a minimum evaluation protocol for fair comparison, and a research gap agenda linked to evidence. These outputs provide researchers and practitioners with a foundation for building reliable, secure, and governable RAG systems in real-world, enterprise, and safety-critical applications.