This thesis explores the verification of factual accuracy in outputs generated by Retrieval-Augmented Generation (RAG) systems. It begins by introducing Large Language Models (LLMs), outlining their historical development, fundamental principles, and the unique characteristics that distinguish RAG systems from standalone LLMs, particularly regarding their use of external databases. The thesis addresses the challenges of evaluating the truthfulness of generated statements, delving into logical frameworks, definitions of truth and falsehood, and current methods and approaches in fact-checking. The issue of "hallucinations" in LLM outputs is critically examined, including how dataset quality influences the factual integrity of generated responses.

The primary objective is to analyze and develop methods for verifying the outputs of RAG systems. The research includes an evaluation of existing fact-checking algorithms, knowledge graphs, statistical and heuristic techniques, and explores the potential of semantic analysis and other LLMs in the verification process. It contrasts manual and automated verification methods, assesses hybrid approaches that combine human oversight with machine-based techniques, and discusses their effectiveness, scalability, and robustness. The thesis also involves implementing proposed verification methods, detailing the underlying technologies and algorithms, and evaluating their performance against predefined objectives.

Expected outcomes include a practical implementation of the developed verification methods and a thorough analysis of the results. The thesis will conclude with recommendations for improving RAG system reliability and propose potential avenues for future research in the verification of AI-generated outputs.