Retrieval-Augmented Generation (RAG) combines the strengths of parametric models with non-parametric memory to improve factual consistency and reduce hallucinations. This survey examines recent architectural innovations, training strategies, and evaluation methodologies in RAG systems published between 2024 and 2026.
RAG remains a promising direction for building trustworthy and up-to-date AI assistants. Open challenges include efficient retrieval at scale and better attribution of generated text to source documents.
References: [1] Lewis et al., 2020; [2] Guu et al., 2020; [3] Izacard & Grave, 2021.