document-understanding

2022.10

false

Document Understanding User Guide

DELIVERY:

Automation Cloud Automation Cloud Public Sector Automation Suite Standalone

Last updated Oct 17, 2024

About Document Manager

UiPath Document Manager is a lightweight web application that allows users to prepare, review and make corrections to datasets required for Training and Evaluation of Document Understanding Machine Learning models. It enables multiple users to perform a variety of operations:

Define and configure the fields to be extracted by an ML model.

Import documents for labeling.

Prelabel documents using a preexisting ML model such as Invoice Extraction or Receipt Extraction provided by UiPath out-of-the-box, or by using a model trained using AI Center.

Label documents.

Export documents in the format expected by the AI Center Training pipelines.

The User Interface

The Document Manager interface contains the following panels:

Management Bar

Displayed at the top of the page in Document Manager.

Enables you to perform multiple operations: navigate in between documents, delete/restore a document, search/filter documents, run AI model predictions, import and export documents.

Here are the items available in the management bar:

Item	Icon	Description
Navigation		Navigate between documents that match the active filter. In between the two arrows, a counter is displayed. It illustrates the number of the current document out of the total number of documents that match the active search/filter.
		Search or filter documents. Filter is also applied when exporting documents. You can also filter by words from a document or by document names.
Delete / Restore	/	Delete or restore a document. Deleted documents can be found under the deleted filter.
		Open Import data dialog box.
		Open Export files dialog box.
Document name and type	n/a	The name of the currently active document and its type. There are three type of documents: Training document Validation document Evaluation document Training and Validation documents are part of training datasets used by Training Pipelines. Evaluation documents are ignored by Training Pipelines and are intended to only be used by Evaluation pipelines in AI Center. These documents are the ones that were marked as evaluation by selecting the Mark this an evaluation set checkbox in the Import data dialog box.
Download		The option is available in the drop-down next to the document name. Click the icon to download a Zip file containing the original document. Besides the original document, all pages converted internally by Document Manager to `.jpeg` images are downloaded as well.
Permanently delete		The option is available in the drop-down next to the document name. Permanently deletes individual files. The `.pdf` and all its `.jpeg` images are deleted from the AI Center dataset and all the metadata is deleted from the database. When clicking the button, a pop-up message appears asking you if you are sure you want to permanently delete the document. Click OK to continue or Cancel to revert to the previous screen.
Session name	n/a	The name of the current session.
Predict		Run AI model predictions and display the results. After configuring Prelabelling, the button is enabled in the management bar. Click it to prelabel the current document. At the moment, using the Predict option with Public Endpoints prelabels only the first 10 pages of a document. This is a known issue and a fix is in the working. Using the Predict option with ML Skills in AI Center, however, does not impose such a limitation.
Settings		Configure OCR and Prelabelling settings or access the How to... panel. See more details below.

Delete and Permanently Delete options

Let's go a little bit deeper in understanding the difference between Delete and Permanently Delete options.

The Delete option deletes the files, but not removing them entirely from your project. The deleted files can still be found under the deleted filter from the Search bar and restored by using the Restore option.
The Permanently Delete option deletes the selected files without any possibility of restoring them.
Observe the use of both options in the below GIF:

Search option

The Search bar is both a text input field and a drop-down.

Search options can be inputted either by writing in the the Search bar or by selecting a filter from the drop-down. There are three main ways of initializing a search:

Using the built-in filters that are available in the Search bar's drop-down. You can choose any of the following filters: train-set, validate-set, train-validate-set, evaluation-set, deleted, labelled, unlabelled.
Note: Please note that for Forms AI only the following built-in filters are available: deleted, labelled, unlabelled.
Using the import batch names. These are also available in the Search bar's drop-down. If added by hand, the format is batch:name, where name is replaced with the name you gave a batch at import time, e.g. batch:invoices1
Using keywords. You have to enter the keyword(s) as free text in the Search bar. The search looks for the keyword(s) in a document's content or the document name.

You can choose to use one or more search options. Every additional option used casts a more specific searching net. Here are some search examples that start off by casting a wide net and slowly progress to a more refined search:

initiating a labelled search returns all the labelled docs in the dataset.
initiating a batch:invoices1 search returns all the docs that are part of the invoices1 batch.
initiating a labelled batch:invoices1 search returns all the labelled docs that are part of the invoices1 batch.
initiating a labelled batch:invoices1 vermont search returns all the labelled docs from the invoices1 batch which contain the inputted keyword, in this case vermont, either in the document name or document content.

The Search bar has a drop-down menu that, when opened, displays the following filters:

train-set - Indicates the number of documents to be used for training the model. Automated action.
validate-set - Indicates the number of documents to be used to validate the model after its training is complete. The split between the train and validate set is targeted to be 80%-20%. Automated action.
train-validate-set - Indicates the number of documents found in both the train-set and validate-set filters. Automated action.
evaluation-set - Indicates the number of documents that had the evaluation set checkbox checked during import and are intended to be used to evaluate the model in the stage of the training pipeline. More information can be found here. Manual action.
deleted - Specifies the number of deleted documents. More information can be found here.
labelled- Specifies the number of docs that have labels. A label is defined by at least one tagged/manually edited field per document.
unlabelled - Specifies the number of docs that don't have labels.
batch:name - Specifies the documents that have been comprised in the same import action.

The allocation of a document to either the train or validate sets is done by the application at import time.

Imported document end up in the evaluation set if the evaluation set checkbox is checked during import.

Settings Menu

The settings button has two available options:

Settings where you can configure the OCR service or Prelabelling
How to... which has the purpose of a help menu

In order to import documents into Document Manager, it is mandatory to configure an OCR service.

The following options are available:

OCR method

The cloud-based options are:

UiPath Document OCR - https://du.uipath.com/ocr;
Google Cloud Vision OCR which has the best language coverage;
Google Cloud Vision OCR for Japanese optimal for reading Japanese documents;
Microsoft Read OCR.

OCR URL

Configuring the OCR requires the OCR service to have a URL. Here are the possible URLs you can use:

public URLs such as https://du.uipath.com/ocr or third-party URLs from Google Vision OCR or Microsoft Read OCR

OCR key

The corresponding API Key for the selected OCR engine. For example, for UiPath Document OCR, you need to use the Document Understanding API Key. Mandatory for Document Manager Cloud and Document Manager On-Prem Online. It is not required for Document Manager On-Prem Air-gapped.

Prelabelling

If you already have a model which can extract some of the fields that need labeling, and there are only a few extra fields that require manual labeling, you can save a lot of time by using Document Manager’s Prelabelling feature.

The following options are available:

Prelabelling URL

Prelabelling requires the ML model has a URL. Here are the possible URLs you can use:

public URLs such as https://du.uipath.com/ie/invoices or https://du.uipath.com/ie/purchase_orders
see the full list of endpoints here
URLs of ML Skills which have been made public in AI Center on-premises or in AI Center Cloud

Prelabelling key

The Document Understanding API Key. Mandatory for Document Manager Cloud and Document Manager On-Prem Online. It is not required for Document Manager On-Prem Air-gapped.

How to...

The How to... option accesses the Document Manager help menu.

Here you can find:

The Document Manager version
The Documentation link leading to this documentation page.
The Labeling Controls section which displays the controls to be used when handling data.
The Document Shortcuts section which displays the shortcuts used to perform various operations such as navigation and UI scaling.
The Configuration section which displays details about the instance configuration as performed during installation.

Column Fields

Column fields have the following options:

Create new column field
Edit field
Expand/collapse column field values

For more details on column fields, visit this section.

Regular Fields

Regular fields have the following options:

Create a new regular field
Edit field

For more details on regular fields, visit this section.

Classification Fields

Classification fields have the following options:

Create a new classification field
Edit field

For more details on classification fields, visit this section.

Document View

For multi-page documents, you can scroll naturally through the pages as in any PDF viewer. To zoom in or out, use Ctrl + mouse scroll.

You can label documents by selecting the word boxes and assigning them to a field by pressing a key. You can also right-click the word box and verify the extracted information.

For more details on how to label documents, visit this page.

When you open a new Document Manager session or when you have an empty filter, certain guidelines are displayed in document view: