OceanBase Data Agent Tops DAB With 90.62% Accuracy

DAB was introduced by UC Berkeley EPIC Data Lab and Hasura PromptQL. Its evaluation covers internet and local life, finance and stocks, biomedicine, intellectual property, enterprise operations, government and public management, and media and entertainment. It also involves multiple databases, including PostgreSQL, MongoDB, SQLite and DuckDB.

Unlike traditional Text-to-SQL evaluations, which mainly test whether AI can convert natural language into SQL, DAB focuses on whether AI can find correct answers from complex, scattered and varied data after entering a real data environment. In a complete task, a Data Agent needs to understand data, select data, plan an analysis path, complete query and computation, and validate results. It tests the full ability from understanding data to obtaining an answer.

As a result, DAB tests not only the underlying model but the combined ability of model, Agent and data system. The model handles understanding and reasoning, the Agent handles planning and execution, and the data system supports data discovery, association, computation and result validation. Whether the three can work together directly affects whether AI can truly put enterprise data to use.

OceanBase's submission was built around this capability. The system uses DataLens to construct data profiles and identify fields and data relationships. It then plans execution paths according to task complexity, completing data selection, filtering, association and computation. After obtaining results, it checks the calculation process and results through evidence tracking and answer validation. When problems are found, it adjusts the plan and revalidates, forming a closed loop of data understanding, planning and execution, and validation and repair.

This explains the weight of the 90.62% result. It does not simply prove that a Chinese model can complete data queries. It validates that a Chinese database and a Chinese model, when combined, can support complex data analysis tasks. With GLM-5.2 as the underlying model, the OceanBase submission ultimately exceeded several Data Agent entries built on overseas models including GPT, Claude Opus and Claude Fable.

For OceanBase, this is also a new validation of its AI capabilities. In the past, databases were mainly responsible for storing, querying and processing data. As AI Agents become new users of data, the problems databases need to solve are changing. AI must not only obtain data but also understand data, associate data, analyze data and verify whether results are reliable. The data foundation in the AI era is therefore beginning to move from handing data to AI toward helping AI put data to use. This is the direction of OceanBase's evolution from a database to an AI data platform.

OceanBase DataPilot is a product direction for AI data analysis scenarios. The DAB evaluation was submitted under the internal codename Scout, and the related technical capabilities will later be incorporated into the DataPilot product. They will further take on data understanding, task planning, analysis execution and result validation, allowing AI to move from calling data to using data to complete tasks.

From database to AI data platform, what changes is not only the product form but also the relationship between databases and AI. In the past, databases stored data, computed it and retrieved it. Now they also need to help AI understand data, organize data and complete analysis. Topping DAB is a validation of OceanBase extending database capabilities further into the AI data analysis stage, and it means the database is moving from an AI data foundation toward an AI data work engine.

Editor's Summary

OceanBase's Data Agent Scout ranked first on the international DAB benchmark with 90.62% accuracy, the first entry to exceed 90%, using OceanBase and GLM-5.2. The result surpassed several GPT- and Claude-based entries and is expected to feed into OceanBase DataPilot. It also signals OceanBase's push from database infrastructure toward an AI data platform.