Datasets
Every transaction belongs to a dataset. Datasets control where the transaction is stored and how it appears in the UI. To target a specific dataset, set dataset_id when creating a transaction via the Multi-Step API.
| Dataset | dataset_id value | Description |
|---|---|---|
| Production | "production" | The default dataset. Transactions appear in the Work Queue for review and production use. |
| Playground | "playground" | For testing and experimentation. Transactions appear in the Playground tab. |
| Memory | "memory" | Used for few-shot learning examples that help improve AI accuracy. |
| Evaluation | "{dataset_uuid}" | User-created datasets for quality evaluation. To find the UUID, open the dataset selector dropdown on the Evaluation tab, choose Manage Evaluation Datasets, and use the Copy ID command from the dataset's context menu. |
When dataset_id is not specified, transactions are created in the Production dataset.
Who can use each dataset
Production needs only the transaction permissions (create, view, review). Playground, Memory and Evaluation are design-time datasets, the project's workspace for trying out its configuration: listing them, opening or changing their transactions, creating transactions in them and copying transactions into them all need the Configure project permission (project.configure) in the project, on top of the usual one. Without it the call is refused with 403 and permission_denied_grant_required, naming the permission. Among the built-in roles, Admin and Designer hold it.
One exception: adding production transactions to Memory (Add to Memory on the Work Queue, or POST .../datasets/memory/copy-transactions with production sources) needs only the System can learn from user permission, since it is how a reviewer teaches the system; Manage memory does not replace it there. Copying Playground or Evaluation transactions into Memory needs Configure project and either Manage memory or System can learn from user.
Visibility applies in every dataset: a caller who sees only its own transactions sees only its own in Playground, Memory and Evaluation too.
Production (Work Queue)
The default. Transactions land in the Work Queue for human review and production processing.
{
"dataset_id": "production"
}
Use this when:
- Processing real documents in a production pipeline
- Documents should appear in the Work Queue for review
- Building an automated document processing integration
Playground
For testing and experimentation without affecting production data.
{
"dataset_id": "playground"
}
Use this when:
- Testing new workflow settings or extraction fields
- Debugging document processing issues
- Running experiments before deploying to production
Memory (Few-Shot Learning)
For adding documents as training examples that improve future AI accuracy.
{
"dataset_id": "memory"
}
Use this when:
- Building a library of correctly-labeled examples
- Improving extraction accuracy through few-shot learning
- Curating training data for specific document types
Whatever is in Memory is what the system learns from, so creating a transaction here, and editing or re-running one, needs the Manage memory or System can learn from user permission in the project as well as the usual one. Without it the call is refused with 403 and permission_denied_memory_edit.
A change to a Memory transaction is made on its draft copy, and production memory changes only when the project is published. When the transaction you address is also used by a published version, the change is made on a new draft copy instead, and the response names it:
| Header | Meaning |
|---|---|
X-Memory-Transaction-Id | The draft copy the change was made on. Use this id for further changes to the sample. |
X-Memory-Forked-From | The transaction id you sent. |
A sample that has been removed from the draft cannot be changed: the call returns 409 with memory_sample_not_in_draft.
Evaluation
User-created datasets for measuring and comparing processing quality. Each evaluation dataset has its own UUID.
{
"dataset_id": "your-evaluation-dataset-uuid"
}
Use this when:
- Measuring extraction, classification, or splitting accuracy against ground truth
- Comparing processing results across different project settings versions
- Running quality benchmarks before deploying settings changes
Open the Evaluation tab in the UI, click the dataset selector dropdown, and choose Manage Evaluation Datasets. Right-click a dataset (or open its context menu) and select Copy ID to copy the dataset UUID to your clipboard. Use this UUID as the dataset_id when submitting transactions via the API.