A BIM-Specific RAG Database for LLM-Based Building Design
Overview
- Current LLMs struggle with BIM editing and generation. They understand IFC terms, but often miss the exact entity definitions, property sets, naming rules, and APIs needed to create valid models.
- They need the right information at the right time, such as BIM best practices, reference components, building standards, regulations, and API documentation.
- Goal: Build a BIM-specific RAG system that provides this information to the LLM through document search or callable tools.
Tasks
- Collect and study BIM data, such as components, properties, standards, regulations, and code documentation.
- Store the data in a suitable database, such as a vector database.
- Develop retrieval tools and connect them to an LLM.
- Compare different retrieval methods.
- Test whether the retrieved information improves LLM-based BIM tasks.
- Open to exploring further directions.
Expected Outcome
- A searchable RAG database for BIM components and documents.
- A pipeline that connects the database to an LLM.
Context
The topic connects to our work on tool use in large language models and to the Show2Instruct project, where an LLM edits IFC models through an MCP-based tool server. The retrieval system built here is can be used to plug into that setup.
Core References
- Nithyanantham, B. K., et al. (2025). “MCP4IFC: IFC-Based Building Design Using Large Language Models.” arXiv preprint arXiv:2511.05533. [web]
- Lewis, P., et al. (2020). “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS.
- IfcOpenShell , the open source IFC toolkit and geometry engine.