Skip to main content
Roberty Studio
IDPBuilder

Identification

Check that the incoming document is the expected kind before extraction in an IDP template: comparison with the reference, minimum similarity, fraud detection and what happens when it fails.

Identification

Identification checks that the incoming document is the expected kind before anything is extracted. It works as a gate: a proof of address sent to an invoice template, for example, can be turned away before spending extraction, post-processing and validations on it.

Identification is off by default. To use it, turn on the Identification enabled switch in the stage.

info

Identification needs the generative AI feature enabled on the workspace; without it, the stage shows as locked. Contact your account manager to have the feature turned on.

How the comparison works​

Identification visually compares the first page of the incoming document, as it arrived and before the pre-processing filters, with the first page of the template's reference document. The document type and the additional criteria add a text description of what is expected to the comparison.

The analysis is made by the model set in Model, always with temperature 0.

Fields​

FieldDefaultDescription
Document type—The expected type, for example "Invoice", "ID card" or "Proof of address".
Additional criteria—Details that help identify the right document, such as the issuer, the layout or fields that must be present.
Minimum similarity80%From 50% to 100%, in steps of 5%. Below this, the document is not recognized as this template's.
Recapture/fraud detectionOnFlags screen photos, reprints, photos of printed sheets and composites.
On failureRejectWhat happens when the document does not pass: Reject or Flag. See below.

When the document passes​

The document passes identification when all three conditions are met:

  1. The model considers it a match for the expected document.
  2. The similarity to the reference is equal to or higher than the Minimum similarity.
  3. With Recapture/fraud detection on, no fraud signal was found.

The stage's result carries the similarity, whether the document was accepted or rejected, the detected type, any fraud signals found and an explanation of the decision. The detected type, the signals and the explanation are written in Portuguese.

Reject or flag​

OptionWhat happens when the document does not pass
RejectThe document is turned away and the pipeline stops: extraction, post-processing and validations do not run.
FlagThe pipeline carries on as usual, and the identification result is recorded as failed, flagging the document for review.

A rejection is a valid pipeline outcome, not an error: the run ends as Finished, with identification failed and the following stages marked as not run. In the Playground, the stage shows Pipeline stopped. A failure of the model itself during the analysis, on the other hand, ends the run as Failed; the document is never accepted by default.

To make sure wrong documents are turned away, add cases to the evaluation sets with documents of another kind and the expected outcome Expected failure.

Without a reference document​

Without a reference document there is no image to compare against: identification classifies the incoming document using only the Document type and the Additional criteria. In that case at least one of the two must be filled in; with both empty, the run ends with an error.

To upload a reference document, see Document and pages.