A chat assistant is only useful when it can act on the system it sits in, and only safe when it acts with no more rights than the person typing. Over the second half of 2024 the Studio assistant gained three tools and one rule: its SQL goes through the same query layer as the application's, so object and row security are injected before anything runs.
Three tools
- Chart a query The model supplies the SQL, the chart type and the axis columns; the platform runs the query and renders the chart. Nine chart types: pie, bar, horizontal bar, line, area, half pie, polar, radar and dot. The tool description is templated with the connection's SQL dialect, and on a SQL error the model is told to rethink the query for that dialect.
- Find an object A dictionary search tool locates tables, reports and other objects by name or description, and a later tool returns an object's statement.
- Read the manual A retrieval tool answers from the documentation index, with Redis Stack search added as an alternative store to the vector database.
One rule
- SQL under the user's security Assistant queries execute through the XML query layer, so the object security and row-level rules that apply to a screen apply to the model.
In the chat
- Slash to pick a tool Typing a slash in the chat input offers the available tools.
- Tables and charts in replies A reply can carry a rendered table or chart, and a developer can open the message log behind any answer.
The tool set has grown considerably since. The security rule has not changed, and it is the reason the assistant could be opened to business users at all.