Materialization in BigQuery
In BigQuery, Live Query can temporarily store the sampled results of eligible AI tools so that the tools don't need to run again each time you interact with downstream tools.
During workflow design, AI tools process a limited sample of the input data. Live Query reuses the sampled results while the AI tool configuration and its input remain unchanged. The interface indicates when preview results are sampled. Downstream previews are based on the same sample.
Full-data AI processing occurs when you run the workflow, not automatically during preview.
Nota
Design-time results are previews based on sampled data. Run the workflow to process the complete dataset.
Workflow Refresh
Workflow Refresh refreshes the workflow's source data and marks materialized preview results as outdated. It does not immediately reprocess those results. The next time Live Query needs a preview, it recalculates the affected results using the refreshed source data.
Recalculate Preview Results
Live Query recalculates a preview when:
You change the tool configuration, prompt, instructions, or selected input column.
An upstream change affects the rows or values sent to the tool.
You change the model, provider, or model version.
You refresh the source data.
You select Workflow Refresh.
The stored preview expires.
Nota
To avoid reprocessing existing AI results during a workflow run, configure the workflow so that populated results bypass the AI tool and only new or empty rows are processed.
Eligible Tools
These BigQuery tools support materialization:
Tool | Materialization Sample |
|---|---|
Classify | 500 rows |
Document Extract | 2 rows |
Score | 500 rows |
Speech to Text | 2 rows |
Text Summary | 500 rows |
Translate | 500 rows |