A multimodal retrieval-augmented generation framework for intelligent campus question answering

Traditional question-answering systems exhibit significant limitations when processing multimodal and heterogeneous information, restricting their effectiveness in complex scenarios. To address this challenge, this paper proposes an intelligent question-answering system based on a multimodal Retrieval-Augmented Generation (RAG) framework, focusing on key technical challenges such as multi-source heterogeneous data processing and cross-modal semantic alignment. This framework integrates three core technologies: BGE-M3 for text embedding, CLIP for cross-modal representation learning, and BGE-Reranker for precise reordering. Through a multi-stage retrieval mechanism and context-aware generation strategy, it effectively processes multimodal data including text, images, and tables, achieving unified cross-modal semantic modeling and efficient retrieval. At the implementation level, we designed and deployed a highly concurrent, scalable intelligent question-answering platform for university faculty and students, validating its stability and responsiveness at Wuhan University of Science and Technology. Experimental and testing results demonstrate that the system outperforms traditional approaches in retrieval accuracy, response speed, and user interaction satisfaction, fully showcasing the feasibility and advantages of the proposed framework. This research provides a referenceable technical pathway for the engineering implementation of multimodal RAG frameworks in intelligent question-answering systems, laying a foundation for future development of multimodal data-driven intelligent interaction systems.