Skip to main content
An extraction can request a human review when a boolean condition you define evaluates to the desired trigger state (typically false). This lets you route edge cases to a reviewer directly in the dashboard — without Automations.

​
Introduction

Retab can flag extractions that need a second look based on conditions you control. Add any boolean in your schema (or computed during inference) and connect it to the human-in-the-loop criteria. When the condition triggers:
  • requires_human_review: true in the projects/extract response (field on the completion object)
  • human_review_flag: “pending” in the stored extraction and in the dashboard

​
Define a condition in your schema

You can designate any boolean field as a human-in-the-loop trigger. For example, define a field like is_document_an_invoice and instruct the model to act as a judge for this condition during inference. If the condition evaluates to the trigger state (e.g., false), the extraction is flagged for review.

​
Example condition

  • Field: is_document_an_invoice: boolean
  • Trigger: When is_document_an_invoice is false
  • Outcome: The extraction requires human review and appears in the dashboard with a pending status

​
API response (projects/extract)

When you call POST /v1/projects/extract/{project_id} (or client.projects.extract in the SDK), the response is a RetabParsedChatCompletion. Use requires_human_review to know if a human review is needed.
(6)
{
  "id": "cmpl_...",
  "object": "chat.completion",
  "model": "auto-small",
  "choices": [
    {
      "index": 0,
      "message": {
        "parsed": { /* your fields, including is_document_an_invoice */ }
      }
    }
  ],
  "requires_human_review": true
}
The dashboard tracks review state via human_review_flag on the stored extraction.

​
Review in the dashboard

  • Where: Project → Monitor -> Files -> Requires review
  • Filter: Use the “Requires review” filter (internally human_review_flag = "pending").
  • Actions:
    • Inspect the parsed output and the document
    • Make corrections to the extracted object as needed
    • Mark the extraction as reviewed
Once reviewed, the extraction moves to human_review_flag = "reviewed" and is removed from the pending queue. That’s it — use boolean conditions to route edge cases to reviewers and keep your pipeline robust.