We are happy to introduce AI Descriptions in the September release.
AI Descriptions are a new type of semantic metadata, at the column-level, for datasets and subject-areas.
This metadata is consumed by the AI model when a question is semantically related to that column and AI description. It is improving significantly the accuracy and consistency of the Agents responses.
When to add an AI Description ?
Add an AI Description when the column name and synonyms do not fully explain an important fact about the column.
This is a complementary metadata that should explain a real specific business meaning or a specific detail the Agent must know to interpret it correctly.
This complementary metadata should explain a real business meaning or a specific detail the Agent needs to interpret the column correctly.
An AI Description belongs to the column and may be used by several Agents. It does not replace the Agent’s supplemental instructions: supplemental instructions define Agent behavior, while AI Descriptions define dataset semantics.
For example, an AI Description can:
- Distinguish similar columns.
- Clarify a column’s grain or unit.
- Explain how its values are formatted.
- Describe its relationship with another column.
- Explain important inclusions or exclusions.
- …etc.
Here two concrete examples:
Example 1: Avoiding duplicate values
The dataset contains the columns Invoice Amount and Invoice Type.
Invoice Type has four possible values: 'Original Invoice', 'Adjustment', 'Credit Memo', 'Consolidated Total'. 'Consolidated Total' already includes the other invoice types. Treat 'Consolidated' Invoice Type as a single filter, it should not be aggregated with the other invoice types.
This context helps the Agent avoid filters that would result in double counting.
Example 2: Understanding value formats
The dataset contains the columns Fiscal Calendar Name and Fiscal Quarter.
The label format depends on Fiscal Calendar Name. 'Corporate Fiscal Calendar' uses YYYYQn, such as 2026Q2; 'Regional Fiscal Calendar' uses Qn-YYYY, such as Q2-2026.
This context helps the Agent use the correct stored value when applying a fiscal-quarter filter.
Without these AI Descriptions, the AI model may choose an incorrect filter or column value and generate an inaccurate query.
Where do I create the AI description ?
For a dataset, go to Inspect > Search > AI Descriptions.
As a quick start, you can generate AI Descriptions automatically by selecting Generate. Make sure AI descriptions add a meaningful information beyond the column name and synonyms.
For a Subject Area, go to Console > Search Index > AI Descriptions.
Now It’s Your Turn
AI Descriptions are most effective when they are selective, precise, and carefully reviewed. Use them where additional business context can materially improve query accuracy, and validate the Agent’s responses with representative questions.
