Text2BIM: Generating Building Models Using a Large Language Model-based Multi-Agent Framework

This repo contains the implementation of the Text2BIM paper: https://www.arxiv.org/abs/2408.08054

The conventional BIM authoring process typically requires designers to master complex and tedious modeling commands in order to materialize their design intentions within BIM authoring tools. This additional cognitive burden complicates the design process and hinders the adoption of BIM and model-based design in the AEC (Architecture, Engineering, and Construction) industry. To facilitate the expression of design intentions more intuitively, we propose Text2BIM, an LLM-based multi-agent framework that can generate 3D building models from natural language instructions. This framework orchestrates multiple LLM agents to collaborate and reason, transforming textual user input into imperative code that invokes the BIM authoring tool's APIs, thereby generating editable BIM models with internal layouts, external envelopes, and semantic information directly in the software. Furthermore, a rule-based model checker is introduced into the agentic workflow, utilizing predefined domain knowledge to guide the LLM agents in resolving issues within the generated models and iteratively improving model quality. Extensive experiments were conducted to compare and analyze the performance of three different LLMs under the proposed framework. The evaluation results demonstrate that our approach can effectively generate high-quality, structurally rational building models that are aligned with the abstract concepts specified by user input. Finally, an interactive software prototype was developed to integrate the framework into the BIM authoring software Vectorworks, showcasing the potential of modeling by chatting.

The tool_agent folder contains the Python implementation of the multi-agent framework. The web_palette folder contains the C++ plugin, which allows the framework to be used via the UI in Vectorworks.

Please note that the web_palette folder is covered by a different license/agreement. All experiments are done in a Windows PC.

Update

• Now supports Vectorworks 2025 and 2024. Please refer to web_palette for more details.
• Now supports the latest LLMs, including Claude-4, o4-mini, and Gemini-2.5-pro. Please ensure you have the corresponding API keys.
• Fixed several bugs in the toolset and improved error reporting.
• Extended the syntax supported by the Python interpreter.
• Enhanced the experiment logging mechanism.

Getting start

0. You need to have Vectorworks 2024 or 2025 and Solibri Office installed in your PC (also need valid license to use both software).
1. Install packages in Vectorworks using powershell. Adapt to your own path and the version you want to use (2024 or 2025).

   cd "C:\Program Files\Vectorworks 2024\Python39\Scripts"; 
   .\pip.exe install -r "C:\Users\xxx\Desktop\Text2BIM\requirements.txt" --no-cache-dir --target="C:\Users\xxx\AppData\Roaming\Nemetschek\Vectorworks\2024\Python Externals"
   

2. Add the project root path in Vectorworks's Tools>Plug-ins>Script Options so that it can be recognized.
3. Config your API keys and PROJECT_ROOT path in tool_agent\agents.py or .env. Config the LLM model names/versions in tool_agent\multi_agents_workflow.py as well as in tool_agent\agents.py if needed.
4. Refer to web_palette\README.md for web palette setup. The plugin can be found in Windows>Palattes>Web Palettes>Text2BIM in Vectorworks.
5. Open tool_agent\solibri_data\LLM-checking-FullExample.smc using Solibri. You can right-click the sample model to update its path to the actual path in your computer, like C:\Users\xxx\Desktop\Text2BIM\tool_agent\Ifc_test_data\sample_model.ifc. Select "Automatic" mode in the pop-up window. Save the updated smc file and close it.
6. Run the prompts in the test_prompt.txt in data\test_env_(version)_new.vwx to replicate the experiments in the paper. The initial experiments (preprint v1) were done in June-July 2024. The latest experiments were done in May 2025. So if possible try to reproduce the experiments using the version of the LLM models available in this time period. For example, in initial experiments we used the gpt-4o-2024-05-13 version.
7. When exporting IFC from Vectorworks, please save the IFC file under tool_agent\Ifc_test_data, i.e. OUTPUT_FOLDER_PATH defined in tool_agent\multi_agents_workflow.py. Make sure the name of the IFC file is the same as the name of the latest created prompt_(uuid..).txt in the same folder, so that Solibri can automatically load the IFC for checking.

Config Gemini

We use Vertex AI and google genai SDK for Gemini setup, if you want to try Gemini. We use the vertex AI version instead of gemini api because it is more stable.

1. Install gcloud cli https://cloud.google.com/sdk/docs/install?
2. Follow instruction to set up cloud project on vertex https://cloud.google.com/vertex-ai/generative-ai/docs/start/quickstarts/quickstart-multimodal
3. Install googel gen AI Python SDK https://cloud.google.com/vertex-ai/generative-ai/docs/sdks/overview (should be already installed via requirements.txt)
4. Config your cloud project name in tool_agent\agents.py

Acknowledgement

The implementation is based on https://github.com/VectorworksDeveloper/SDKExamples and https://github.com/huggingface/transformers

Citation

@misc{du2024text2bimgeneratingbuildingmodels,
      title={Text2BIM: Generating Building Models Using a Large Language Model-based Multi-Agent Framework}, 
      author={Changyu Du, Sebastian Esser, Stavros Nousias and André Borrmann},
      year={2024},
      eprint={2408.08054},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2408.08054}, 
}