Retrieval-Augmented Generation: Architectures, Advances, and Real-World Applications

Abstract—Large Language Models (LLMs) have demonstratedremarkable capabilities in natural language understanding andgeneration; however, they suffer from hallucination—the generation of factually incorrect yet fluent text—and are limitedto knowledge encoded at training time. Retrieval-AugmentedGeneration (RAG) addresses both limitations by grounding modelresponses in dynamically retrieved external knowledge. This paper provides a comprehensive survey of RAG as it stood in 2024,covering its mathematical foundations, architectural evolutionfrom Naive RAG to Advanced and Modular RAG, retrievalmechanisms including dense and sparse retrieval and hybridfusion, and the integration of long-context LLMs with retrievalpipelines. We benchmark RAG variants on open-domain questionanswering (Natural Questions, TriviaQA) and summarisationtasks, demonstrating consistent gains over vanilla LLM baselines.Applications in healthcare, legal reasoning, enterprise knowledgemanagement, and code generation are reviewed. Finally, wediscuss open challenges including retrieval latency, faithfulnessevaluation, multi-hop reasoning, and the tension between longcontext models and retrieval-based approaches. Our analysisestablishes RAG as the dominant paradigm for knowledgeintensive NLP tasks in 2024.Index Terms—retrieval-augmented generation; RAG; largelanguage models; dense retrieval; hallucination; knowledgegrounding; question answering; vector databases; hybrid retrieval; agentic RAG