Exécution de notre outil de synthèse sur un nouvel article

Résumé : Large language models are susceptible to problems such as lagging parameter knowledge updates, insufficient alignment of textual and graphical evidence, and incomplete factual basis for answers in domain-specific question answering. To address these issues, a domain-consistent multimodal retrieval enhancement generation framework, DC-MRAG, is proposed. This framework comprises four modules: domain anchor index, hierarchical cross-modal retrieval, evidence consistency fusion, and uncertainty-aware generation. It achieves unified representation of textual and graphical evidence through domain concept constraints, utilizes multi-factor scoring for candidate recall and re-ranking, compresses context based on evidence support, and controls generation risk through confidence calibration and selective secondary retrieval. Under a unified example experimental setting, the proposed method achieves an exact match rate of 84.9%, a recall@5 rate of 91.4%, an answer fidelity of 94.2%, and a phantom rate reduction to 4.6%. Compared to RAMQA, the exact match rate is improved by 4.2 percentage points, the phantom rate is relatively reduced by 41.03%, context tokens are reduced by 26.99%, P95 response latency is reduced by 23.43%, and throughput is increased by 28.85%. The results show that the collaborative design of domain indexing, cross-modal retrieval, evidence fusion, and generative calibration can improve the accuracy, evidence reliability, and online service efficiency of knowledge-based question answering.