OceanBase Data Agent Tops International DAB Benchmark With 90.62% Accuracy

The benchmark was launched jointly by UC Berkeley EPIC Data Lab and Hasura PromptQL. Its evaluations span internet and local-life services, finance and equities, biomedicine, intellectual property, enterprise operations, government and public administration, and media and entertainment, and cover databases including PostgreSQL, MongoDB, SQLite and DuckDB.

Unlike conventional Text-to-SQL tests, which mainly check whether an AI system can turn natural language into SQL, DAB asks whether an AI can locate the correct answer inside a real data environment made up of complex, scattered and inconsistently structured sources. A complete task requires the Data Agent to understand the data, select the relevant data, plan an analysis path, run queries and computations, and verify the results.

That design means the benchmark measures a combination of model, agent and data system rather than the underlying model alone. The model handles understanding and reasoning, the agent handles planning and execution, and the data system supports data discovery, relational computation and result checking.

OceanBase's submission, code-named Scout internally, is built around that loop. It uses DataLens to build a data profile and identify fields and their relationships, then plans an execution path according to task complexity and carries out data selection, filtering, joining and computation. After producing a result it traces evidence and checks the answer, revising the plan and re-verifying when problems appear, forming a closed cycle of data understanding, planning and execution, and verification and repair.

The scheme was built on the Chinese large model GLM-5.2 and finished ahead of several Data Agent entries built on overseas models, including GPT, Claude Opus and Claude Fable, according to the report.

For OceanBase, the ranking extends work it has been doing as AI agents become a new class of data user. Databases have traditionally handled storage, queries and processing; agents now need to understand data, relate it, analyze it, and check whether results are reliable. OceanBase DataPilot is the company's product direction for AI data analysis, and the capabilities tested under the Scout codename will be folded into DataPilot, covering data understanding, task planning, analysis execution and result verification.

The article was provided by Ant and republished by QbitAI with authorization, with views attributed to the original author.

Editor's Summary

OceanBase's Data Agent entry, submitted under the internal codename Scout, ranked first on the Data Agent Benchmark with 90.62 percent accuracy, the first submission to exceed 90 percent. The result was produced with the Chinese large model GLM-5.2 and beat several entries built on GPT, Claude Opus and Claude Fable. OceanBase says the underlying capabilities will be integrated into its DataPilot product for AI data analysis.