Retrieval-Augmented Generation: A Systematic Survey of Architectures, Chunking Strategies, and Evaluation

Large language models generate fluent text, but they are bounded by the knowledge frozen in their parameters at training time, they hallucinate facts, and they cannot inspect private or fast-changing information. Retrieval-Augmented Generation (RAG) addresses these limits by coupling a language model with an external retrieval system: relevant documents are fetched at inference time and supplied as context, grounding generation in verifiable evidence. This paper presents a systematic survey of the RAG literature. It organizes the field into a unified pipeline of indexing, retrieval, augmentation, and generation; compares dense, sparse, and hybrid retrieval architectures; examines chunking strategies and their effect on retrieval quality; reviews evaluation methodology, including benchmark datasets, retrieval metrics, and generation metrics; and discusses open challenges such as retrieval failure modes, context-length trade-offs, evaluation contamination, and the economics of maintaining a retrieval corpus. The survey concludes that RAG has shifted the central design problem of language-model systems from model scale alone toward corpus quality, retrieval precision, and evidence-aware generation.