RAMASC: A retrieval-augmented multi-agent framework for automated structural calculation

초록

Large language models (LLMs) have recently gained traction in construction engineering for interpreting technical documents, extracting domain-specific information, and supporting natural language interfaces. To mitigate hallucinations and improve factual reliability, retrieval-augmented generation (RAG) has been widely adopted, enhancing the factual consistency of generated responses using external reference documents. However, current RAG-based LLM applications still struggle with performing accurate structural calculations, especially when mathematical reasoning and code compliance are required. To address these challenges, this paper proposes a Retrieval-Augmented Generation-based Multi-Agent System for Structural Calculations (RAMASC), which integrates semantic retrieval, executable code synthesis, and multi-agent debate for reliable automated structural computation. RAMASC consists of (1) a vector database for retrieving relevant design code provisions and structural equations, (2) a prompt-based LLM module that generates executable Python code with interpretable reasoning, and (3) a multi-agent system in which multiple LLM agents independently generate and cross-verify responses to reduce errors and improve consistency. The framework was evaluated in two structurally different verification contexts: precast concrete floor panel design and inverted T-type retaining wall verification. RAMASC achieved an average accuracy of 95.5% for the slab cases and 91.3% for the retaining wall cases. These results suggest that the proposed framework can support reliable code-based structural calculations across heterogeneous structural verification tasks.

키워드