Roll Your Own Virtual Ontology with Claude Code

This project demonstrates an approach to semantic data access: using a "virtual ontology" layer that enables natural language REPLs (like Claude Code) to directly translate business questions into SQL. By combining traditional ontology concepts with modern LLM capabilities, we can query existing databases without the overhead of formal semantic systems.

An alternative data access pattern:

Traditional: Raw Data → ETL → Reports → Insights

Virtual Ontology: Raw Data → Semantic Layer → Natural Language Querying → Actionable Insights

Clone the repo, open a Claude Code session, load @sys_prompt.md in chat.

The virtual ontology approach:

1. Keep data in place - Query existing SQL databases directly
2. Define business concepts - Lightweight ontology (TBox/RBox) maps concepts to schema
3. Leverage agent capabilities - Natural language REPLs handle the translation
4. Learn from usage - Capture successful patterns for reuse

Through testing with real manufacturing data:

• Identified significant improvement opportunities through systematic analysis
• 86% query success rate on first attempt, 98% with single refinement
• Handles temporal patterns, cascading failures, financial calculations

Questions like these translate effectively to SQL:

• "Which equipment is the bottleneck on each line?"
• "What's the financial impact of material jams?"
• "Show cascade failures from upstream equipment"
• "Find quality issues specific to morning shifts"

• Intent capture with every query
• Successful pattern extraction and reuse
• Gradual improvement through usage

• Handles 36,000+ record datasets efficiently
• Supports complex aggregations and window functions
• Integrates well with python (visualization + advanced analysis)

This repository includes a working implementation using Manufacturing Execution System (MES) data that demonstrates the approach:

1. Capacity Analysis: Compare best vs current performance
2. Bottleneck Detection: Identify constraining equipment
3. Quality Investigation: Find patterns in defect rates
4. Downtime Impact: Calculate financial cost of stoppages
5. Cascade Analysis: Trace upstream/downstream effects

Defines business concepts, relationships, and rules:

• Classes: Equipment, Products, Events, Downtime Reasons
• Relationships: upstream/downstream, belongs-to, produces
• Business Rules: OEE calculation, cascade patterns, quality impact

Maps ontological concepts to database structure:

• Column mappings with business names
• Data types and constraints
• Indexed fields for performance

Intent-aware query execution with pattern learning:

• Captures business intent (≤140 chars)
• Logs successful patterns
• Builds reusable query library

Traditional Semantic Systems Virtual Ontology Approach

Requires ETL to RDF/OWL	Works with existing SQL databases
SPARQL expertise needed	Natural language queries
Complex triple stores	Simple SQL execution
Static ontology definitions	Evolving pattern library
Slow iteration cycles	Rapid exploration
High implementation cost	Low barrier to entry

This project explores the intersection of semantic technologies and modern LLM capabilities. We welcome:

• Use case implementations in new domains
• Ontology improvements and extensions
• Pattern learning enhancements
• Performance optimizations
• Documentation and tutorials

If you use this virtual ontology approach in your work:

[To be determined]

---

This project demonstrates that combining traditional ontology concepts with modern agent-based tools can create practical solutions for semantic data access.