ruc-datalab/EvoOntology

EvoOntology: A Self-Evolving Ontology Layer for Data Agents ⚙️ EvoOntology插件为Claude Code/Codex 建立&进化本体层

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entity relationships—and eliminates the need for costly, manually maintained static semantic layers that quickly become outdated.

How it works

It implements a self-evolving ontology layer that acts as trainable agent state. The system follows a specific lifecycle:

  1. Build: It derives candidate concepts from workloads and verifies them against raw data to create an initial version.
  2. Use: Data agents query this layer on-demand via MCP (Model Context Protocol) tools, rather than having the entire ontology injected into the prompt.
  3. Evolve: An evolution agent analyzes interaction trajectories to diagnose errors and propose localized updates.
  4. Evaluate: New candidates are compared against the parent version using the same data and budget to ensure reproducible improvement.
  5. Publish: Only validated improvements are published as new versions.

The ontology consists of three layers: a Content Layer (a semantic graph of terms, mappings, and constraints), a Schema Layer (rules for representation), and a Tool Layer (MCP tools for browsing and resolving semantics).

Who it’s for

Developers building data agents that need to operate over heterogeneous data sources and those who want to reduce the agent's reliance on repeated data exploration and manual semantic mapping.

Highlights

  • MCP-compatible: Seamlessly integrates as a plugin for agents like Claude Code and Codex.
  • Self-Evolving: Continuously refines its knowledge base based on actual execution behavior.
  • Workload-Grounded: Only commits semantic objects after verification against underlying data.
  • Gated Versioning: Ensures that ontology updates only occur when they provide a measurable performance gain.

Related

  • Project
  • Project
  • Project
  • Project
  • Project