Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large Language Models (LLMs) have achieved remarkable progress in natural language understanding and generation. Despite their success, hallucination—where models generate fluent but factually incorrect or unsupported information—remains a major challenge that limits their reliability in real-world applications. This study aims to develop and evaluate a hybrid framework that improves factual consistency by combining Retrieval-Augmented Generation (RAG) with Milvus-based vector search and Low-Rank Adaptation (LoRA) fine-tuning. The proposed system retrieves semantically relevant documents using Sentence-BERT embeddings, integrates the retrieved evidence through structured prompt engineering, and employs parameter-efficient fine-tuning to guide the model toward grounded and evidence-based responses. The hybrid model is evaluated against a baseline LLaMA-2-7B model using Semantic Similarity, Hallucination Rate, Fact Score, and Accuracy metrics. Experimental results show a significant reduction in hallucination rate (51.00% to 20.00%) and notable improvements in factual accuracy (24.30% to 60.67%). These results demonstrate that integrating vector-based retrieval with efficient fine-tuning provides a practical and scalable approach for enhancing the trustworthiness of LLM systems.