Guidelines for adding new documents

INFO

This page describes the flow for requesting, preparing, developing and deploying new data extraction models, also called SmartDocs. Its goal is to guide how documents should be sent and to ensure that extraction and classification keep high accuracy levels through continuous improvement, data governance and performance monitoring.

How to request a new SmartDoc

The lifecycle of a new extraction model starts with a customer request.

To ensure traceability, proper prioritization and alignment between the teams involved, new SmartDocs must be requested formally through one of these channels:

Send your request by email directly to the support team.

Help Desk

Open a ticket on the Help Desk platform to register your request.

Sales team

Discuss your needs directly with the Sales team.

Preparing and sending the documents

To develop a highly accurate model, we need to assess the document type and receive a representative sample. The better the quality, variety and clarity of the samples, the better we can identify patterns, layout variations and the fields that are critical for extraction.

Document categories

Structured

Semi-structured

Unstructured

Documents with a fixed pattern and defined regions where the data is placed. In some cases, positions may shift proportionally: when one region moves, the others move in the same proportion.

Common examples include identity documents, bank slips (boletos) and forms with a standardized layout.

Submission requirements

Send the sample

Send 10 to 20 different images, in good quality, legible and without cuts, folds or significant obstructions. Whenever possible, the samples should cover the main variations expected in production.

Map the desired fields

Clearly indicate which data must be extracted. You can do this in text or by marking the fields on a sample image, as shown in Figure 1.

Include special cases

Send examples with relevant variations, such as dynamic tables, optional fields, layout changes or sections that may appear in different positions, as illustrated in Figures 2 and 3.

WARNING

Samples that are illegible, cropped, folded, low resolution or lacking layout diversity can reduce the model's accuracy and increase the time needed for analysis, adjustment and validation.

Visual examples

Figure 1 — Marking the fields to be read

Figure 2 — Dynamic table 1

Figure 3 — Dynamic table 2

Our ModelOps flow

Once the request is registered, the process follows a technical ModelOps pipeline. This pipeline organizes how documents come in, supports developing or refining models, sets validation criteria and ensures that models are released to production with proper governance.

Triage and clustering

We organize the incoming flow and automatically group similar documents. Files are processed in a controlled, ephemeral environment to ensure compliance, privacy and traceability during the analysis.

Development and refinement in SmartDoc Studio

We map new layouts or fine-tune existing models when extraction confidence is low (a reduced score or the need to adapt to new layout variations).

Quality assurance with Ground Truth

We build a canonical dataset through manual review, called Ground Truth, to validate the model's accuracy. The automatic extraction is compared against this dataset before the model is promoted to production.

Deployment and governance

The approved model is promoted to production. The data is kept temporarily for training audit purposes and then discarded in accordance with our sensitive data governance policies.

CHECK

A model only goes to production after its accuracy has been validated against the canonical dataset defined during the Ground Truth stage.

Timelines and endpoints

New SmartDocs are integrated into the extraction endpoints that match the document type. The right endpoint depends on how structured the document is, how predictable its layout is and the nature of the data to be returned.

Document type	Estimated timeline	Recommended solution
Structured	10 business days*	`content-extraction`
Semi-structured	10 business days	`content-extraction` or `multipage-content-extraction`
Unstructured, mostly text	10 business days	`generative-content-extraction`
Unstructured with visual elements (such as charts and complex tables)	10 business days	`agentic-content-extraction`

NOTE

* For structured models, delivery may happen within the first 5 business days, depending on the field specification, the quality of the samples sent and the complexity of the variations identified.

INFO

This service focuses on raw data extraction. To apply business rules, custom validations, calculations or automated flows after extraction, we recommend Connect, the solution dedicated to these integrations.

Guidelines for adding new documents

How to request a new SmartDoc#

Preparing and sending the documents#

Document categories#

Submission requirements#

Visual examples#

Our ModelOps flow#

Timelines and endpoints#