Text-to-Siheyuan: LLM-based Multi-Agent Procedural Generation System of Vernacular Architecture in Beijing Keywords: Large Language Model, Multi-Agent System, Procedural Content Generation, Generative Design, Vernacular Architecture Abstract: Rule-Based Procedural Content Generation (PCG) methods have been instrumental in exploring the mathematical principles underlying traditional Chinese vernacular architecture (Wang et al. 2020). However, they often require detailed architectural knowledge to understand and use the procedural systems, creating a high barrier to usage. This research presents an LLM-based multi-agent procedural generation system(Wu et al. 2023) for traditional Chinese vernacular architecture, enhancing efficiency in analyzing design requirements and facilitating accurate decision-making while maintaining cultural fidelity. Siheyuan, a quintessential example of traditional Chinese architecture, is chosen as a case study (Ma 1999). Our methodology comprises three steps: 1. Component-Based Disassembly of Traditional Design Principles: We divided Siheyuan planning principles into four components—Architecture, Entrance, Veranda, and Wall. Each component serves as a foundation for our procedural generation system, managed by specialized agents. 2. Collaborative Multi-Agent Procedural Modeling: Four GPT-4-based agents (Courtyard Planner, Courtyard Divider, Depth-Width Planner, and Component Manager) were employed, each with a specific role in synthesizing parameters. Users input design requirements through natural language, and the bot, guided by questions, generates a regularized fixed-format procedural DSL (Domain-Specific Language), ensuring the Siheyuan layout aligns with traditional design principles. 3. Real-Time Interaction and Visualization in 3D Modeling Platforms: The bot's API was integrated with Grasshopper, converting design parameters into our Siheyuan DSL represented in JSON data. The JSON data is processed by the Grasshopper Python API to generate a 3D Siheyuan model, reflecting traditional Chinese architectural structures. This approach offers designers an intuitive and efficient method for generating vernacular architecture models, lowers the professional threshold for public engagement, and enhances participation in cultural heritage preservation and appreciation for the historical value of traditional architecture.