Skip to content
DocuExtract

← All docs

Review queue

The safety net under verbatim grounding. When the engine isn't confident, it asks rather than guesses.

When fields land here

A field lands in the review queue when at least one of these is true:

  • Confidence below the template threshold. Default threshold is 0.75. Anything lower is flagged.
  • Grounding failed.The LLM proposed a value but the engine couldn't trace it back to source text in the document. The engine drops the value (verbatim-grounding invariant) and flags the field for human review with no value attached.
  • Multi-pass disagreement.On the Premium + multi-pass API tier, if two independent LLM passes disagree and a third tiebreaker doesn't resolve it, the field flags for review with all three candidate values in the audit log.

Reviewing

On /dashboard/review each row shows:

  • The field name + type + confidence
  • The document filename
  • What the engine extracted (if anything)
  • The verbatim source text + page number

Two actions per row:

  • Approve— accept the engine's value as-is. The field clears the review flag; the value is unchanged in downstream exports.
  • Correct— supply the right value. The original is preserved in the audit trail; the corrected value becomes the effective value for the field. Optional notes field for context (e.g. "OCR misread the suffix").

The audit trail

Every action is an append-only audit event:

  • Original value (before correction)
  • Corrected value (same as original for approve-as-is)
  • Actor (your user ID)
  • Timestamp + optional notes

Past corrections never disappear — required for compliance workflows that need to reconstruct how a field reached its final state.

Tuning the threshold

If you're seeing too many false positives in review (good values being flagged), lower the template's confidence_threshold. If you're seeing wrong values slip through downstream, raise it.

Default 0.75 is a reasonable mid-point. Production workflows with high stakes typically use 0.85+; high-volume low-stakes bulk extraction uses 0.55–0.65.

From the API

See the API reference for the review endpoints. GET /v1/review + POST /v1/review/{field_id} with action: approve | correct.