Agent-Based Framework for Engineering Document Comprehension with Large Language Models

Abstract Engineering documents, such as design rule books and technical requirements, play a critical role in helping engineers ensure that their designs meet specified standards. Quickly retrieving and reasoning over such documents is essential for efficient engineering workflows. Recent advances in large language models (LLMs) and generative AI make it possible to automatically interpret and reason over complex engineering documents. However, engineering covers many distinct and highly specialized subdomains. Because general purpose LLMs such as GPT or Gemini are not trained or fine-tuned on these domains, their performance on domain-specific reasoning remains limited. It is impractical to curate dedicated datasets for every engineering subdomain, which would be required for both fine-tuning and building reliable retrieval-augmented generation (RAG) frameworks. Therefore, it is crucial to develop a general approach that can fully leverage the capabilities of existing large models and improve performance across diverse engineering subdomains. To address this, we propose an agent-based framework that leverages existing LLM capabilities through specialized agents coordinated by a routing mechanism, achieving a scalable and fully automated workflow. We evaluate our framework on DesignQA as a representative case study and show that it consistently outperforms baseline methods. Moreover, the results highlight its potential to generalize across diverse engineering requirements. In future work, we plan to extend this framework to additional engineering domains and explore more advanced agent strategies.