Human-in-the-loop review queue

Uncertain extractions and unmatched documents pause for a person, then the pipeline resumes from the approved step.

Documents in

  • Any document type with a schema

How it runs

  1. Review gates are enabled

    Review can be turned on for text extraction, classification, schema matching, field extraction, index values and summary.

  2. Pipeline pauses

    When a gate applies, the job pauses with a review status and the document shows that it is awaiting review.

  3. Reviewer works the queue

    The reviewer edits text, classifications, field values or the summary in the portal and sees the extraction confidence.

  4. Approval resumes processing

    Approving enqueues a resume message, and the pipeline continues from the next stage rather than starting over.

  5. Corrections are reused

    Approved corrections are kept and supplied as context to later extractions with the same schema.

What you get

  • Uncertain values held for a person instead of passing silently
  • Documents with no schema match surfaced for a decision
  • Clear status showing which documents are waiting and where
  • Reviewer corrections that improve later runs for the same schema

Where review fits

Some gates skip automatically when the system is highly confident, for example schema matching and field extraction, while others always pause when review is enabled. A document that matches no schema is flagged and can be assigned a schema or sent for schema creation. When a new schema is created, flagged documents can be reprocessed automatically.

Integration points

Review state is visible through the API and job status, and completed results are delivered via API and webhooks. Access to review actions follows role-based permissions, and approvals are recorded for audit.

See this workflow on your documents