Retrieval-Augmented Generation for Large Language Models: Foundations, Evolution, and Practical Applications.

Recently, Large Language Models (LLMs) have become one of the most impactful advances in artificial intelligence, especially for natural language tasks. However, LLMs still show important limitations, particularly when they need up-to-date information, factual accuracy, or clear and traceable reasoning. Retrieval-Augmented Generation (RAG) has emerged as a promising solution to these problems. It combines a language model with an external retrieval module, allowing the system to access relevant documents during generation. This thesis provides an overview of how LLMs work, their main weaknesses, and why retrieval is necessary. Then it presents the foundations of RAG, including its architecture, training methods, and decoding strategies. The thesis also examines how the RAG paradigm has evolved, describes several techniques designed to improve the reliability and accuracy of RAG systems, and discusses some real-world applications of RAG. Finally, it reviews the main challenges, limitations, and future research directions. Overall, this work shows the growing importance of RAG in making modern language models more reliable, precise, and useful.