EvoOntology: A Self-Evolving Ontology Layer for Data Agents

A runtime knowledge layer for data agents, agents that answer natural-language requests over tables, files and databases, from the ruc-datalab group at Renmin University of China (Meiduo Chong, Shaolei Zhang, Ju Fan and Xiaoyong Du). The authors call the problem the agent–data gap: the data sits outside the agent, which can reach it only through generic tools that expose things like column names and file paths. Existing systems either let the agent explore raw sources or paste a hand-built semantic layer into the prompt. EvoOntology instead serves the ontology as an MCP server with schema, content and tool layers that the agent queries at runtime, retrieving only the terms and mappings the current step needs. A builder agent constructs the first ontology by probing the data sources, and a self-evolution loop analyzes agent trajectories to find deficiencies and attribute them to the ontology's schema, content or tools, proposes targeted edits, and keeps each edit only if it passes a paired evaluation on held-out validation tasks with the same backbone model. The repository ships plugins that let Claude Code and Codex build, evolve and explore ontology layers over a user's data. It is MIT-licensed and passed 540 GitHub stars within about two weeks; the paper was posted on September 14 and the repository created the next day.

Six backbones (GPT-5.5, GPT-5.6 Sol, Claude Sonnet 5, Claude Opus 4.8, DeepSeek-V4-Flash and Qwen3.5-Flash) are each run in the same ReAct scaffold, and EvoOntology improves every one of them on all three benchmarks. On DDR-Bench's 10-K scenario, trajectory-wise accuracy rises by 17.8 points on average, from +4.8 for Qwen3.5-Flash to +26.7 for GPT-5.5, while injecting the same content as a static prompt often hurts (−15.0 on Claude Sonnet 5). Averaged over four backbones it reaches 89.5, against 75.8 for a ReAct agent with episodic memory and 69.5 without. On BIRD text-to-SQL it adds 7.4 points of execution accuracy and 8.6 of valid efficiency score on average, lifting Claude Opus 4.8 from 67.5 to 78.3 execution accuracy, and on InsightBench it gains 1.9 points on average. The builder's initial ontology gives the larger step and self-evolution adds a further gain on each benchmark (+12.3 then +7.7 points on DDR-Bench, +5.1 then +3.7 on BIRD). The abstract mentions four backbones; the main result tables report six.