EvoOntology Gives Data Agents a Self-Evolving Semantic Layer

Introduction

For a language model operating on real-world data, the hardest part is often not writing SQL or calling a file tool. It is understanding what the data represents. Tables, files, and databases may use inconsistent names, layouts, and business terminology. When an agent can only inspect these sources through generic tools, it must spend many steps discovering structure and meaning. A manually prepared semantic layer can help, but it is expensive to build, difficult to scale, and unlikely to fit every agent or dataset.

EvoOntology addresses this agent–data gap with an ontology layer that agents can actively query and that can change over time.

How it works

• An MCP-based interface: The ontology is exposed as an MCP server, allowing an agent to request semantic information during execution instead of relying on a static prompt injection.
• Three complementary layers: EvoOntology separates the ontology into a schema layer, a content layer, and a tool layer. The first describes how data is organized, the second adds meaning grounded in the data, and the third supports interaction with the resulting knowledge.
• Autonomous construction: A builder agent creates a data-specific ontology directly from available sources, reducing the amount of manual semantic engineering required.
• Controlled evolution: Interaction history can trigger attribution-guided, typed edits. Changes are not accepted unconditionally: the system compares the revised ontology with the previous version and uses a backbone-conditional paired evaluation before committing an update.

Why it matters

The paper evaluates the approach on three established data-agent benchmarks with four LLM backbones. Its abstract reports consistent improvements over strong baselines and existing semantic-layer methods. The result suggests that an ontology for data agents can be more than a static catalog or prompt supplement: it can function as a runtime coordination layer between data discovery, semantic interpretation, and tool use.

This design also reframes agent improvement. Instead of asking a model to remember every detail about a heterogeneous data estate, the system gives it a structured service that can be queried, checked, and refined. The MCP packaging could make that service easier to connect to different agent runtimes, while the separation between schema, content, and tools provides a clearer place to diagnose failures.

Self-evolution nevertheless raises important reliability questions. An update may introduce a mistaken relationship or reinforce a model-specific interpretation. Evaluation must therefore detect not only obvious errors but also subtle regressions across tasks and backbones. EvoOntology’s contribution is to place ontology editing inside an attributable and comparative acceptance loop, rather than allowing unrestricted automatic rewriting. The released implementation offers a basis for testing this idea on broader data environments.

Source: Hugging Face Daily Papers