Setting up human review

AI extraction can be wrong, so Droqify lets a person confirm results at the points that matter. This guide explains how the review queue behaves and how to get the most from it.

Where review fits in the pipeline

A document moves through stages: text extraction, classification, schema matching, field extraction, then indexing and summary. In the human-in-the-loop (HITL) setup, the pipeline can pause after a stage so a reviewer can check or correct the output. Review steps include text extraction, schema matching and field extraction.

While paused, the document carries a status that starts with awaiting_hitl_, followed by the step, for example awaiting_hitl_schema_matching. This is the signal to show the document in a review queue.

Open the review queue

Paused work appears in the portal’s review screen. It lists jobs waiting for a person and opens each one at the step where it paused. Reviewers need the matching permission, so assign it through roles rather than sharing a high-privilege account.

Review a document

1. Open the next item from the queue.
2. Check the output for the paused step. At text extraction you can fix extracted text. At schema matching you can change the classification or the schema mapping. At field extraction you can edit extracted values and compare them with the source.
3. Edit anything that is wrong. Edits are saved to the document.
4. Approve the step.

Approving does not simply mark the item done. It records the review action in the document’s review history and sends the job back into the queue so processing resumes at the next stage. Edited values are kept, not overwritten by a later stage.

Schema pipelines resume separately

Schema generation and multi-type analysis are their own pipelines. If they pause for review, approving resumes that same schema pipeline and does not push the source document through its normal processing again. You do not need to manage this, but it explains why a source document may not change status while its schema is under review.

Flagged documents

A document that matches no schema is flagged. The flagged list offers four actions: assign a schema, validate a schema against the document, reprocess, or dismiss. If the document is already paused at a review step, assigning a schema updates the matches and clears the flag without advancing the pipeline; you then approve as usual.

Corrections

Reviewer edits are recorded as corrections. Administrators can browse and delete them. Corrections are used as examples within your own tenant to guide later extraction.

Choosing what to review

Review everything while you are new to a document type, then narrow. Common patterns are to review only low-confidence results, only required fields, or only the first documents from a new sender. Track how often reviewers change a value; that rate, measured on your own documents, is the honest signal for when to relax review.