How We Prototyped an LLM Interface for Geometrid API

The Case for AI in the Built Environment

The architecture, engineering, and construction (AEC) industry has long been characterized by complex data structures, specialized terminology, and information silos. Building Information Modeling (BIM) has shaped how we design and construct buildings, but the technical expertise required to access and manipulate this data remains a significant barrier to widespread adoption.

Large Language Models offer a compelling solution to this challenge. By acting as intelligent intermediaries between humans and complex technical systems, LLMs can democratize access to building data, streamline workflows, and reduce the technical knowledge required to leverage powerful BIM capabilities. The potential applications span the entire building lifecycle, from design and construction to operations and maintenance.

However, for LLMs to provide real value in construction, they need access to structured, reliable building data. This is where platforms like Geometrid, with our robust APIs, create new possibilities for innovation.

Leveraging Geometrid API

When we set out to explore the possibilities of construction data management, we developed a proof of concept that combines Geometrid's API with language models. Our implementation demonstrates how natural language understanding can be used to translate conversational inputs into precise API operations, showcasing potential new ways to interact with BIM and field data.

Our Approach

At the core of our experimental system is a command router that interprets natural language inputs and maps them to specific Geometrid API endpoints. The proof of concept shows how users could potentially filter assets, update statuses, and manage records through conversational commands.

The system employs several complementary techniques: semantic parsing to extract structured information from natural language, confidence thresholds to ensure mapping accuracy, and Levenshtein distance calculations to find the closest matches between user terminology and standardized IFC properties.

Additionally, the router leverages context-aware processing that incorporates project-specific information from configuration files to improve command interpretation.

Key Features

The prototype reveals four distinct capabilities that leverage Geometrid's API through language model integration. Each demonstrates a different dimension of how conversational interfaces can enhance traditional building information data access and management.

Multi-Model Intelligence

Our implementation supports switching between different LLMs (e.g. Gemma2, Llama3, Phi4) to optimize for different tasks. This flexibility demonstrates how we could select the most appropriate model for each type of query or command in a production environment.

1Your prompt (or '/quit' to exit):
2'Switch to llama3.2'
3INFO | Commands: switch_model
4INFO | Switched: llama3.2

By matching task requirements with model strengths, the system can leverage Phi4 for technical precision in property matching while using Llama3 for creative problem-solving and workflow orchestration. This adaptability ensures optimal performance across diverse construction queries and commands.

Natural Language Processing

The proof of concept demonstrates how natural language commands can be translated into precise API calls, handling complex parameter extraction and request formatting automatically. Users can issue intuitive queries like "progress for precast elements on level five" or "assign status delivered to aluminum facade panels," which the system translates into appropriate Geometrid API operations.

1Your prompt (or '/quit' to exit):
2'Last week progress for my curtain wall on south elevation'
3INFO | Commands: filter_assets, report_progress
4INFO | Filter: 1025 matching assets
5INFO | Progress: 0 Produced, 64 Shipped, 12 Delivered, 84 Installed

This approach dramatically simplifies interaction with construction data by eliminating the need for technical query syntax or deep knowledge of API structures while maintaining the full power of the underlying system.

Record Data Management

The solution shows how record operations can be streamlined through conversational commands, enabling users to create, retrieve, and assign metadata to assets with simple natural language inputs. For example, users can create records from external files, bulk assign attributes to multiple elements, or retrieve specific project data - all without navigating complex interfaces.

1Your prompt (or '/quit' to exit):
2'Mark all steel brackets on fifth floor for replacement'
3INFO | Commands: filter_assets, assign_records
4INFO | Filter: 72 matching assets
5INFO | Records: Replace (Defects) assigned

This demonstrates how metadata management workflows could be transformed from technical tasks to intuitive conversations, potentially reducing training requirements and accelerating data operations across construction projects.

Entity Mapping Capabilities

One of the most powerful aspects we explored was the ability to semantically map terminology from various sources to Geometrid entities processed from standardized IFC properties. The system can analyze files with custom field names and match them to the appropriate properties in the Geometrid system, maintaining high confidence thresholds to ensure accuracy.

1Your prompt (or '/quit' to exit):
2'Link cost data from bill-of-quantities.csv'
3INFO | Commands: create_records, map_entities, assign_records
4INFO | Records: 2285 CostItems created
5INFO | Mapping: 2161 entities mapped
6INFO | Records: 2161 records assigned

This capability shows how the integration could potentially bridge the gap between organization-specific terminology and industry standards.

Technical Foundation

Our experimentation builds on Geometrid's comprehensive API architecture:

• RESTful Design: Standard endpoints with predictable patterns ensure reliable data access and command execution
• Authentication Framework: Secure organizational context preservation with token-based access control
• Workflow Management: Structured stage tracking provides context for progress filtering and status updates
• Flexible Record System: Adaptable metadata framework supports diverse project information needs
• Comprehensive Data Model: Rich IFC property mapping enables precise filtering and semantic matching

This well-designed API infrastructure provided the essential foundation for our natural language interface to connect meaningfully with building information systems.

Potential Impact

This approach demonstrates how interaction with construction data could be transformed:

• Broader Accessibility: Reduces technical knowledge barriers to BIM information, making it accessible to all team members
• Intuitive Interaction: Enables users to query complex IFC properties using everyday language rather than technical syntax
• Semantic Alignment: Bridges the gap between organization-specific terminology and industry-standard properties
• Workflow Acceleration: Simplifies multi-step processes into single conversational commands for all team members

Our proof of concept shows how Geometrid's API can be extended to create more intuitive ways to interact with construction data, demonstrating how powerful BIM operations could be made accessible through conversational interaction in future implementations.

The Future of LLMs in the Built Environment

Looking beyond our proof of concept, the integration of LLMs with construction platforms like Geometrid points to several transformative possibilities for the industry:

• Contextual Understanding: Future systems will likely develop deeper understanding of building context—comprehending not just the technical properties of elements but their functional roles, relationships, and importance within the overall building system.
• Cross-platform Intelligence: As APIs become more standardized across the construction technology ecosystem, LLMs could serve as universal translators between previously disconnected platforms—bringing together information from design, construction, and operations into cohesive knowledge graphs.
• Predictive Assistance: By analyzing patterns in historical project data accessed through APIs, AI systems could provide predictive guidance—anticipating coordination issues, suggesting optimal sequencing, or identifying potential quality concerns before they manifest.
• Knowledge Democratization: Perhaps most significantly, these integrations will continue to democratize access to construction knowledge. Field personnel, project managers, and executives without deep technical expertise will gain the ability to query complex building information systems through natural conversation.

The path forward requires thoughtful collaboration between construction technology providers, AI developers, and industry practitioners. APIs like Geometrid's provide the essential foundation, structured data access and command capabilities, while LLMs offer the intuitive interface.

Together, they represent a powerful opportunity to address some of construction's most persistent challenges: information fragmentation, knowledge silos, and the technical barriers that have historically limited BIM adoption. This proof of concept demonstrates that this future isn't distant speculation, the fundamental components already exist.

The next step is bringing these technologies together in robust, production-ready implementations that can deliver on the promise of truly intelligent building information management.

Interested in exploring similar capabilities for your building data? Contact our team to discuss how we can help you develop a solution for your specific needs.