Language Models and Design Theory for Efficient Engineering Knowledge Retrieval

Abstract

Large Language Models (LLMs) are increasingly utilized as a tool for connecting engineers to design knowledge because of the scale and speed with which they automate information retrieval tasks during early-stage design. However, if engineering documentation is input to LLMs without preserving design-specific structures of information, this risks unexplainable and inefficient knowledge retrieval. In this work, we incorporate design theory into the process by which design knowledge is represented for storage and retrieval by LLMs. We construct design-inspired data representations of engineering knowledge with the aim of improving their explainability and performance. We demonstrate these methods on an industry case study of the design of high voltage power transmission structures. We present an implementation integrating LLMs into a design theory-based framework for knowledge representation. We validate the system experimentally with human subject matter experts and conduct a time complexity analysis of the search efficiency of the design theory-based representations in comparison to baseline methods. We conclude that design theory-based representations result in improved computational search efficiency, but that establishing ground truth in engineering design for validation purposes is challenging.