Retrieval-Augmented Generation for Knowledge-Intensive Language Applications

Abstract

Retrieval-augmented generation (RAG) couples a parametric language model with a non-parametric retrieval system, allowing factual grounding to be supplied at inference time rather than baked into model weights. Since the original RAG formulation in 2020, the technique has become the dominant production pattern for question answering, enterprise search, and knowledge assistants. This paper provides a structured survey of RAG: its theoretical motivation, core architectures, advances in retrieval (dense, sparse, hybrid), reranking, query rewriting, and graph-augmented variants. We analyse evaluation methodologies, deployment patterns, and open challenges including factual faithfulness, multi-hop reasoning, and long-context interaction. We argue that RAG is best understood as a system design pattern rather than a single algorithm, and that engineering decisions in chunking, indexing, and orchestration often dominate model-side choices.