Multilingual graph retrieval-augmented generation for product design using design knowledge
Hongbin Zhang, Tao Wang, Zhongyu Liang, Zhenghao Huang, Chong Chen, Lianglun Cheng · Journal of Engineering Design · 2025
Product design integrates knowledge from multilingual sources (e.g. manuals, patents, news) across life cycle stages. However, fragmented and language-specific representations hinder designers' ability to unify domain concepts and leverage cross-lingual insights for knowledge construction and recommendation. In this study, we propose a multilingual design knowledge graph-guided retrieval-augmented generation framework (MDKG-RAG), which can reduce the multilingual design knowledge extraction cost while promoting the domain design knowledge recommendation. Firstly, the multilingual design knowledge is dynamically optimised and extracted from the multilingual design-relevant corpus by LLMs for the design knowledge graph construction module (MDKGC). Next, the multilingual graph retrieval-augmented generation (MGRAG) assembles the constructed design knowledge graph and the collected intelligence database into user query analysis. The integration process can enhance LLM's capability to better understand professional terms by adaptively retrieving key multilingual texts with entity paths. Then, these retrieved multilingual texts are independently encoded with collaborative LLMs in MGRAG, which aims to output the final query answers by generating, aggregating, and validating multiple query answers with multilingual prompts. Finally, experiments on two question-answering datasets, including multilingual ship outfitting design and long document retrieval, demonstrate MDKG-RAG's effectiveness by acquiring the maximum answer similarity of 32.4% and the most context recall of 35.1%.