Exporting Data

Pipelines supports exporting any dataset from the Data Vault — whether it holds agent run results (trajectories, outputs, and judge verdicts) or imported reference data. The export flow differs depending on whether the dataset is pipeline-backed or imported.

Export formats

Exporting pipeline-backed datasets

Pipeline-backed datasets support async export with progress tracking. This handles large datasets with many runs or file attachments.

1. Open a dataset in the Data Vault.
2. Click Export in the toolbar.
3. In the export dialog:
   • Select the format (CSV, JSON, or ZIP).
   • Review the column list. Metadata columns (Run ID, Status) and all field columns are included by default.
   • Optionally deselect columns to exclude them.
   • For structured fields (judge verdicts, ratings, scores), expand the column to select/deselect individual object keys.
   • Optionally rename columns by editing the column header text.
4. Click Export.
5. For large exports, a floating progress indicator appears in the bottom-right corner of the screen:
   • Processing — the export is being generated.
   • Export ready — click Download to save the file.
   • Failed — an error occurred. The message is shown in the indicator.

Exporting with files

For pipelines with file upload fields, select ZIP as the export format. The archive contains:

• A CSV or JSON file with the run data.
• A files/ directory containing all uploaded files, organized by run ID.
• File paths in the data file reference the corresponding files in the archive.

Column controls

When exporting structured fields (like judge verdicts or scores), you can control which sub-keys are included:

• Select/deselect keys — choose which object properties to include in the export.
• Rename keys — provide custom column headers for the exported file.
• Reorder columns — drag columns to change their order in the export.

Exporting imported datasets

Imported datasets support synchronous export — the file downloads directly in your browser.

1. Open an imported dataset in the Data Vault.
2. Click Export in the toolbar.
3. In the export dialog:
   • Select the format (CSV, JSON, or ZIP).
   • Review and optionally adjust column selection, renames, and ordering.
4. Click Export. The file downloads immediately.

Datetime normalization

Datetime fields are automatically normalized to ISO 8601 format in exports, regardless of the display format used in the UI.

Exporting from the Data Explorer

The Data Explorer is the per-pipeline run table on a pipeline's detail page. It supports the same export options as pipeline-backed datasets. See Tasks for details on the Data Explorer.

API export

Datasets can also be exported programmatically, with support for format selection, column filtering, and column renaming.