Retrieval-Augmented Generation Systems

Abstract Modern large language models are fluent but not factual. They hallucinate, they cannot quote sources, and they know nothing about events that occurred after their training cut-off. Retrieval-Augmented Generation, commonly abbreviated as RAG, pairs a generative model with an external knowledge store so that answers are grounded in evidence fetched at query time rather than memorised in the model's parameters. This study examines RAG systems through a descriptive-analytical lens, synthesising academic literature, vendor documentation, and reported benchmarks. Using a comparative approach, the research evaluates the dominant retrieval strategies (sparse, dense, hybrid, late-interaction), the common index structures (Flat, IVF-PQ, HNSW), and the four broad architectural paradigms identified in recent surveys (Naive, Advanced, Modular, and Graph-based RAG) across accuracy, latency, scalability, fault tolerance, and operational complexity. Six predominant challenges are identified: factual hallucination, retrieval relevance, latency and cost, stale indexes, security-including prompt injection through retrieved content-and evaluation difficulty. Practical mitigations drawn from literature, including hybrid retrieval, cross-encoder reranking, citation verification, incremental indexing, instruction hardening, and model-graded evaluation, are presented. Findings indicate that no single configuration dominates; the appropriate design is determined by the application's tolerance for latency, its freshness requirements, and the cost model under which it operates. The paper provides structured guidance for practitioners assembling a RAG pipeline.