Skip to main content
A dataset tests a workflow against real inputs. It is a folder of examples, each with an input and an optional expected output. The folder is the manifest; there is no separate config file.

​
Archive layout

  • input.json is the full run input. Use { "$file": "input/Invoice.pdf" } wherever the input should receive a file.
  • Files under input/ and expected/ are included only when referenced from input.json or expected.json.
  • expected.json is the ground truth an evaluator compares against. Leave it out for examples you only want to run, not score.
Example folder names must match [a-z0-9][a-z0-9-_]*.

​
Build it

You can hand-author the folders, or create scalar-only examples from the CLI:
For examples with file inputs, use the archive layout above: place files under examples/<name>/input/, reference them from input.json, then push the dataset. That keeps original filenames and mirrors the import/export format.

​
Push and pull

The dataset round-trips. Push your local folder, or pull the live dataset to edit it:
push defaults to append; replace swaps the whole dataset and asks you to confirm. The archive layout is identical in both directions, so a single-example export re-imports cleanly when you need to move one case between environments.

​
Next

Score the dataset in Evaluate a workflow.