OceanBase DataPilot (hereinafter referred to as DataPilot) is an intelligent data analytics platform that primarily uses dialogue as its entry point. You can organize data sources, semantic assets, knowledge, tools, and models into Agents, and then use natural language dialogue to generate reusable data applications such as reports, dashboards, large-scale displays, multi-dimensional tables, and workflows, which can be published for authorized users to use.
DataPilot targets two types of users: data engineers, who are responsible for building, validating, and publishing Agents; and data analysts, who subscribe to and use published Agents and applications. The workspaces and feature entrances available to these two roles differ after login. For detailed descriptions, see Roles and navigation paths.
Product positioning
DataPilot leverages three fundamental objects to deliver all its capabilities. Understanding the relationship between these three is key to grasping the full picture of the product.
Workspace: The governance boundary where members, data sources, Agents, and access policies collectively take effect. After switching workspaces, the visible members, data sources, and Agents change accordingly.
Agent: An intelligent work unit composed of models, data scopes, semantic and knowledge assets, Skills, MCP tools, and data applications. Engineers configure it in the workspace, validate it through real conversations, and then publish it for authorized users to use.
Data Application: A result object that is published and authorized along with an Agent, including reports, dashboards, large-scale displays, multi-dimensional tables, workflows, etc.
Note
An engineer's current draft does not automatically become the version seen by an analyst. Only after a successful publish operation should the new state be considered available.
Overall architecture
DataPilot is divided into five layers from bottom to top. The responsibilities of each layer and the corresponding documentation are shown in the following table.
Level |
Main Components |
Responsibilities |
References |
|---|---|---|---|
| Data Access Layer | Data source connection, data table discovery, and data table preview | Securely connect external databases or data services to the workspace and confirm the fields and samples that the current identity can view. The data table preview returns at most 10 rows. | Overview of Data Source Management |
| Semantic layer | Business Objects, Attributes, Relationships, Dimension Values and Aliases, Metrics, and Filter Conditions | Organize physical tables into business language so that the agent understands "what business objects are, how they are related, and what can be computed from them." | Semantic Layer Overview |
| Agent runtime layer | Model binding, data scope, knowledge, skill, MCP tools, and behavioral rules | Organize the resources and execution boundaries that can be used for a request and produce a verifiable response. | Agent Overview |
| Application and Distribution Layer | Ten types of data applications, publishing, subscription, sharing, and exporting | Consolidate dialogue results into reusable assets and distribute them to users based on their permissions. | Start conversation analysis |
| Governance and Observability Layer | Data Permission Policies, Licenses, Monitors, Audits, System Logs, and Task Center | Constraints define the accessibility scope and provide four types of facts—health, operations, service output, and long-running tasks—when issues arise. | Data Permission Model, Monitoring Center, and Audit Logs |
Data permissions continuously take effect between the semantic layer and the Agent runtime layer: adding a data table to an Agent's data scope does not bypass the workspace data policy; the sandbox restricts the execution boundaries of code-based nodes.
Data applications generated via dialogue
DataPilot currently provides ten types of data application workspaces, all of which are generated via Agent dialogue. After generation, they are refined through dialogue iteration or workspace editing based on their type.
Data Applications |
Main uses |
References |
|---|---|---|
| report | Generative or orchestrated reports, version and snapshot management, with support for sharing and exporting | Report Overview |
| Dashboard | Configure queries and charts, support filtering controls, and enable prediction capabilities | Create Dashboard |
| Large Screen | Custom canvas for binding data and materials, used for demonstrations and duty displays | Create Data Dashboard |
| Generative data applications | Generate interactive applications from requirement descriptions and maintain submitted versions. | No special topic available |
| Multi-dimensional table | Workheets support filtering, sorting, and grouping. They also support AI columns, lineage, and result fixation. | Multidimensional Table Basics |
| Notebook | Orchestrate SQL, Python, Markdown, charts, and input controls, run queries, and save snapshots. | Notebook |
| Workflow | Node-based orchestration and testing. The new version takes effect upon releasing the Agent, which then initiates scheduling. | Workflow Overview |
| Spatio-temporal analysis | Build and publish map analysis based on real data | Spatio-temporal analysis |
| Document classifier | Create a classification dataset, train a model, perform batch inference, and export the results. | Document classifier |
| Data Extraction Factory | Configure extraction templates to batch process CSV, XML, image, and PDF files, then store the results in the database or write them to the knowledge base. | Extraction Factory |