Review extracted documents
Open a document from a finished job and read what the pipeline extracted — fields, tables, figures, and selections, each with confidence — and check a value against where it was read on the page.
A job produces documents — one file may split into several — and each document carries the data the pipeline read from it. The Documents view is where you open a single document and read that data: the fields it found, how confident it is in each, and where on the page each value came from.
Anyone confirming a result — an operator spot-checking output or a developer verifying what a pipeline returns before wiring it to a destination. Reading a document does not change it.
1. Open a document
Open a document from its job on the Jobs screen, or from the Documents list. The view is a split layout: the document itself on one side, the data extracted from it on the other.
2. Read the fields and confidence
The extracted values sit alongside the document, each field paired with a confidence score. Confidence is the pipeline's own read of how sure it is — a low score is the signal to check the value, not proof it is wrong. See Extract for what the stage produces and how confidence is assigned.
A processed document: the source on the left with an overlay marking where each value was read, and the extracted fields on the right, each with its confidence. The results split into Fields, Tables, Figures, and Selections, with a raw view.
3. Check a value against the page
Selecting or hovering a field highlights where that value was read on the document — a bounding box on the page. This is the fastest way to confirm a value came from the right place, rather than a plausible-looking value from the wrong part of the page.
4. Tables, figures, and selections
Beyond plain fields, a document's data is grouped by kind — Fields, Tables, Figures, and Selections — with a raw view for the underlying structured output. Tables keep their rows and columns; selections capture checkboxes and choices. Read each group in its own tab.
5. Send low-confidence documents to review
A document that tripped a rule or came back low-confidence is flagged as needing review and is worked in the Reviews workspace — claim it, correct the value, and submit. The Correct a document in review how-to covers that loop; Exceptions & human review explains when a document is routed there in the first place.
To pull the same extracted data programmatically instead of reading it here, retrieve the job over the API — see the job-status endpoint in Submit a job.