Exécution de notre outil de synthèse sur un nouvel article

Résumé : Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLM) by grounding responses in retrieved context, addressing hallucination concerns in knowledge-intensive applications. However, evaluating RAG-based LLMs remains challenging due to complex interactions between retrieval and generation components. This work presents a comprehensive evaluation framework combining RAGAS LLM-based metrics with classical information retrieval measures. We evaluate a RAG pipeline using Qwen 3 14B on the Neural Bridge RAG Dataset 12000, comparing RAG and non-RAG grounded response generation. To address same-model evaluation bias, we introduce cross-model evaluation using Gemma 3 12B, demonstrating that RAG benefits are robust across evaluation models (93-109% improvement in answer correctness). The number of the retrieval chunks proved to be a highly relevant hyperparameter to improve efficiency while maintaining quality. Classical metrics including chunk precision, chunk accuracy and embedding similarity corroborate RAGAS findings while providing computational efficiency.