Natural language BIM search
GQL generation accuracy
Real-time project insights
Deployed on AWS-native stack
The challenge
Simplifying access to complex BIM data
Engineers and architects spent hours navigating technical BIM files to locate specific details. This slowed down decision-making, delayed project delivery, and diverted focus from high-value work.
Key challenges
Manual search through complex BIMs
Delayed decisions and project timelines
Need for natural language interaction with technical data
The solution
GenAI Copilot
Built on Amazon Bedrock using fine-tuned Mistral 7B
Converts natural language to Graph Query Language (GQL)
Claude 3.5 Sonnet used for synthetic training data and response tuning
RAG Framework
Custom Retrieval-Augmented Generation pipeline
Technical documents stored in Amazon S3
Embedded via Amazon OpenSearch Service for semantic search
Implementation approach
1
Model Stack
Amazon Bedrock (Mistral 7B, Claude 3.5 Sonnet)
Amazon CloudWatch for monitoring
Amazon SageMaker for model hosting
2
Data Pipeline
Amazon OpenSearch for vector embeddings and retrieval
Amazon Textract for OCR and table extraction
Amazon S3 for document storage
3
Integration Layer
AWS Lambda for orchestration
Amazon API Gateway for secure access
IAM roles for access control and security
The impact
Business Outcomes
Instant BIM insights
Reduced manual effort
Faster project decisions
Technical Gains
High GQL accuracy
Domain-aligned responses
Robust RAG architecture
User Experience
Natural language interface
Broad adoption across roles
Scalable across use cases
Looking ahead
Next Steps
Expand to new BIM formats
Integrate with CAD and planning tools
Enable voice-based queries
Future Capabilities
Real-time design validation
Predictive compliance checks
Cross-project knowledge graphs
AI Evolution
Continuous fine-tuning
Feedback-driven improvements
Enterprise-wide GenAI rollout
