EvoOntology: 데이터 에이전트를 위한 자가 진화형 온톨로지 레이어
이 논문은 다양한 데이터 소스를 다루는 Data agents가 겪는 agent-data gap 문제를 해결하기 위해 self-evolving ontology layer인 EvoOntology를 제안합니다. EvoOntology는 ontology를 schema, content, tool layer로 구성된 MCP server로 캡슐화하여 agent가 런타임에 능동적으로 상호작용할 수 있게 합니다. 이를 위해 ontology를 자율적으로 구축하는 builder agent와 지속적으로 ontology를 개선하는 self-evolution loop를 도입합니다.
Data agents aim to fulfill natural-language instructions over heterogeneous data, including tables, files, and databases. However, data agents face a challenging agent-data gap: heterogeneous data resides outside the agent, while the agent can access it (e.g., column names and file paths) only through generic tools. Existing approaches either let agents directly explore raw data sources or inject manually constructed semantic layers into prompts. However, neither scales well to large heterogeneous data sources nor adapts to different agent behaviors. In this paper, we introduce EvoOntology, a self-evolving ontology layer for data agents. EvoOntology encapsulates the ontology as an MCP server comprising a schema layer, a content layer, and a tool layer, enabling agents to actively query and interact with the ontology at runtime. To this end, we introduce a builder agent for autonomous ontology construction and a self-evolution loop that continuously refines the ontology through attribut