A Review of Retrieval-Augmented Generation Technology

Abstract

Retrieval-augmented generation has emerged as a core technological paradigm for addressing the bottlenecks of hallucinations and knowledge lag in large language models. However, many existing reviews focus on a single technical branch or vertical application scenario, making only scattered references to hardware, evaluation methods, and cross-industry empirical evidence, and lacking a systematic, end-to-end integration. This paper conducts research based on a total of 115 papers, comprising foundational literature from 1998–2019 and core RAG literature from 2020–2026, systematically cataloging end-to-end technologies and supporting solutions, establishing a quantitative hardware comparison table and comparing 11 categories of open-source and commercial APIs, constructing a two-tier, four-level standardized evaluation framework, and compiling empirical evidence and implementation challenges across eight industries from 2024 to 2026. Based on this, the paper identifies four major structural contradictions—the retrieval–creation trade-off, the geometric–semantic misalignment as a symmetry problem between representation space and semantic structure, the autonomy–reliability paradox, and evaluation blind spots—as a unified analytical framework for the five major technological strands. Finally, this paper proposes four research directions for practical implementation—differentiable joint optimization, hybrid geometric space learning, interpretable causal reasoning, and multidimensional diagnostic evaluation—providing a systematic reference for both theoretical research on RAG and its deployment in the private sector.