Recently, a paper titled “DOKE-RAG: bridging expert knowledge and LLMs via multi-modal graph retrieval for deep vertical domains” was published in the prestigious computer science journal Expert Systems with Applications. The paper’s first author is Zhang Dinglun, a 2025 graduate of SCUT’s Elite English-Taught Civil Engineering program at the School of Civil Engineering and Transportation, currently a master’s student supervised by Professor Chen Taicong. This work builds on Zhang’s university-level outstanding undergraduate thesis advised by Prof. Chen.

Addressing the complexity of civil engineering curricula and the susceptibility of traditional large language models (LLMs) to errors in domain-specific Q&A tasks, the paper proposes a novel multimodal, graph-based retrieval-augmented generation framework named DOKE-RAG. It integrates diverse materials—text, audio, equations, and figures—into a structured knowledge graph and employs a hierarchical LLM-validated entity alignment (Hi-LVEA) algorithm, outperforming existing mainstream models in specialized structural mechanics Q&A tasks.

Furthermore, Zhang’sarticle on his practical experience with AI-assisted scientific illustration attracted over 100,000 reads online. The content has been compiled and officially published asa practical booktitled A Rapid Practical Guide to Nano Banana-Based Scientific Illustration.