RAG-EM: Retrieval-Augmented Generation for Electromagnetic System Design

As large language models (LLMs) gain broader adoption, incorporating domain knowledge via proprietary data is of the utmost importance. However, data such as platform design guides (PDGs) have heavy restrictions in the design flow, meaning that the data must remain on-premise. Thus, using existing language models over the Internet is infeasible without compromising the data. This paper examines using local LLMs with domain knowledge, especially PDGs, with relevant design information and custom tool interfaces. In order to provide the LLM with domain-relevant knowledge, we use modified retrieval augmented generation (RAG). For domain-relevant knowledge, we provide the LLM with numerous PDGs and Python scripts that enable the control of an electromagnetic tool for PCB design. We demonstrate the ability of the model to perform domain-specific question/answering (QA) and generate code for manipulating 3D EM structures such as striplines and vias. We show the advantage of the on-prem RAG-based approach over finetuning by introducing new tasks/instructions.