ruc-datalab/EvoOntology

What it solves

EvoOntology addresses the "agent-data gap" where AI agents struggle to understand the semantics of raw data sources like tables, files, and databases. It solves the problem of semantic uncertainty—where agents must guess metric definitions or entity relationships—and eliminates the need for costly, hand-authored static semantic layers that quickly become outdated as data and workloads change.

How it works

The system treats the ontology layer as a trainable state that evolves over time. It consists of three interconnected layers:

• Content Layer: A typed semantic graph containing terms, mappings, constraints, and evidence.
• Schema Layer: Defines the rules and boundaries for how the ontology is represented.
• Tool Layer: Exposes the ontology to agents via Model Context Protocol (MCP) tools, allowing agents to retrieve only the specific semantics they need on demand rather than processing a full ontology in their context window.

The lifecycle follows a loop of building an initial ontology from workload evidence, using it in agent tasks, diagnosing interaction trajectories to propose updates, and evaluating those updates against a parent version before publishing a new version.

Who it’s for

Developers building data agents that need to operate over heterogeneous and complex data sources without requiring manual semantic mapping or constant expert maintenance.

Highlights

• Self-Evolving: Automatically refines the ontology layer based on observed execution behavior and interaction history.
• MCP-Compatible: Integrates as a universal plugin for agents like Claude Code and Codex.
• Grounded Construction: Semantic objects are only committed after being verified against the underlying raw data.
• Gated Versioning: New ontology versions are only published if they show reproducible improvement over the previous version.