Answers are grounded in this paper and cite their sources.
Large Language Models demonstrate strong generative capabilities but remain constrained by static training data, susceptibility to hallucination, and an inability to access current or domain-specific information at inference time. RetrievalAugmented Generation (RAG) addresses these limitations by coupling language models with external retrieval mechanisms that dynamically supply relevant knowledge during generation. In the years since its introduction, RAG has expanded from a single architecture into a broad ecosystem of retrieval modalities, architectural variants, and deployment patterns that is increasingly difficult to navigate. This survey provides a unified treatment of that landscape, organized around five retrieval paradigms: sparse lexical retrieval, structured data retrieval, dense semantic retrieval, graph-enhanced retrieval, and agentic retrieval. For each paradigm, we trace the evolution from foundational concepts to current practice, examining the core techniques, architectural patterns, and design trade-offs that practitioners face. Cutting across these paradigms, we examine two disciplines that shape every RAG system: context management, which governs how retrieved information is selected, compressed, and presented to the generator, and evaluation, which measures retrieval quality, generation faithfulness, and end-to-end system performance. To make techniques comparable across paradigms, we ground each in a shared vocabulary for the stages of a RAG pipeline and consolidate representative works along five recurring dimensions: retrieval stage, generation stage, system properties, trust and governance, and adaptability. Our analysis reveals that significant challenges persist. Hallucination is reduced but not eliminated, and retrieval augmentation can mask it from standard detectors. Retrieval quality remains highly sensitive to domain and query complexity. Effective use of long contexts lags behind the rapid growth of context window sizes. Automated evaluation metrics often diverge from real-world user satisfaction. We discuss these open problems alongside promising research directions, including multi-modal and federated architectures, cost-efficient graph systems, and standardized governance frameworks. Together, this survey offers a structured road map for researchers and practitioners navigating this rapidly evolving field.
No takes yet. Share an insight, caveat, or question.
Anand et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: