Abstract

Large Language Models (LLMs) generate fluent answers, but they depend mainly on their training data. As a result they can give wrong or outdated responses, and they may lack the specific information a question needs. Retraining a model again and again to fix this takes a great deal of time. Retrieval-Augmented Generation (RAG) addresses the problem by retrieving relevant information from external knowledge sources, such as documents, databases and the web, before a response is generated. This paper presents a literature survey and problem understanding for RAG. It explains the retrieve–augment–generate workflow, states the problem and six objectives, and relates RAG to current technology trends: real-time information, enterprise AI, agentic AI, cost-effective scalability and reduced hallucination. Eight studies from 2020 to 2026 are compared on methodology, key findings and limitations. The analysis reveals that the quality of information retrieval, as well as the effectiveness, accuracy and assessment, still remain unstudied topics. Four research gaps are identified: high computational resources, irrelevant or incomplete retrieval, limited effectiveness in specialised domains, and the need for better evaluation of factuality, relevance and reliability.

Keywords

How to Cite This Article

Rights