Etudiant : IZERIA MOHAMED

Filière : Master Web Intelligence et Sciences des Données (WISD)

Encadrant : Pr. EL FAZAZY KHALID

Annèe : 2025

Résumé : This research internship explores the evaluation of Retrieval-Augmented Generation (RAG) systems powered by Large Language Models (LLMs), with a focus on improving the faithfulness and factual consistency of generated outputs. Traditional LLMs are limited by static knowledge and hallucination risks. RAG addresses this by combining LLMs with external document retrieval, but evaluating such hybrid systems remains challenging. This work investigates RAG evaluation through the RAGAS framework, which automatically generates tests and knowledge graphs from context chunks. We analyze the limitations of RAGAS-generated graphs and propose an enhanced pipeline that integrates semantic validation using the YAGO 4.5 ontology. A novel bidirectional SPARQL-based algorithm is introduced to extract semantic paths between entities. Each knowledge graph triple is compared against YAGO-derived paths using an LLM, which assigns a semantic similarity score and explanation. This allows the detection and correction of incorrect triples, improving the quality and explainability of the graph. The result is a modular, explainable, and semantically enriched RAG system for more robust natural language reasoning.