A Review of Retrieval-Augmented Generation: Architectural Evolution, Key Breakthroughs, and Vertical-Domain Practices

Abstract

Large language models (LLMs) rely on massive pretraining based on the Transformer architecture and show significant emergent effects in general tasks. However, the nature of their statically parameterized knowledge causes vertical-domain applications to face three core bottlenecks: knowledge hallucination, delayed timeliness, and loss in long contexts. Retrieval-Augmented Generation (RAG) realizes the decoupling of computational logic and factual memory through the paradigm of “retrieve first and generate afterward”. It can dynamically inject domain knowledge without fine-tuning and has become a mainstream technical path for the implementation of trustworthy artificial intelligence (AI). This study systematically reviews the evolution of RAG technology from 2024 to 2026, from Naive RAG to Advanced RAG, Modular RAG, and graph-enhanced RAG. It analyzes in depth the breakthrough progress of four key technologies, namely hybrid retrieval, multimodal document parsing, generation control, and lightweight deployment. Combined with typical application cases in complex equipment manufacturing, energy and power, public services, and other fields, it explains the engineering value of RAG. Finally, it discusses the evaluation system, existing limitations, and future trends, providing a reference for academic research and practical implementation of RAG technology.