Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large Language Models have achieved very good results in conversational question answering. However, their efficiency over technical and research documents remains limited due to certain hallucinations, a lack of source grounding, and scarce explainability. Retrieval-Augmented Generation (RAG) mitigates some of these issues due to its incorporation of external knowledge, but most existing RAG systems employ fixed strategies, with little transparency about how retrieved evidence supports generated responses. These limitations reduce trust and usability within high-stakes technical domains. This paper presents an explainable RAG based conversational system, put forward for question answering over technical research documents. The proposed system introduces a query-intent-aware adaptive retrieval mechanism that dynamically adjusts retrieval parameters based on the inferred intent of the user query, allowing for more precise and contextually relevant document retrieval. To ensure answer verifiability, proposed technique leverages citation-grounded response generation explicitly linking the generated outputs to their supporting document segments. An explainability layer is integrated that exposes evidence retrieved and generation rationale for reductions of unsupported claims and improvements in user trust. The system is implemented using an open-source large language model, sentence-level semantic embeddings, and a vector-based similarity search framework-ensuring reproducibility and practical deployment. We conduct experiments with proposed techniques on a dataset of technical research papers, with comparisons against a vanilla LLM model as well as a traditional RAG model. The experimental findings underline that our proposed model performs better w.r.t relevance, citation, and hallucinated answers, and the latency is acceptable, hence emphasizing the effectiveness of adaptive retrieval as well as explanation techniques incorporated in our proposed model, which make proposed a viable answer for technical documents leveraging the benefits of RAG-based conversational systems.