Document Intelligence custom neural model
Important
- Document Intelligence public preview releases provide early access to features that are in active development. Features, approaches, and processes may change, prior to General Availability (GA), based on user feedback.
- The public preview version of Document Intelligence client libraries default to REST API version 2024-07-31-preview.
- Public preview version 2024-07-31-preview is currently only available in the following Azure regions. Note that the custom generative (document field extraction) model in AI Studio is only available in North Central US region:
- East US
- West US2
- West Europe
- North Central US
This content applies to: v4.0 (preview) | Previous versions:
v3.1 (GA)
v3.0 (GA)
This content applies to: v3.1 (GA) | Latest version:
v4.0 (preview) | Previous versions:
v3.0
This content applies to: v3.0 (GA) | Latest versions:
v4.0 (preview)
v3.1
Custom neural document models or neural models are a deep learned model type that combines layout and language features to accurately extract labeled fields from documents. The base custom neural model is trained on various document types that makes it suitable to be trained for extracting fields from structured and semi-structured documents. Custom neural models are available in the v3.0 and later models The table below lists common document types for each category:
Documents | Examples |
---|---|
Structured | surveys, questionnaires |
Semi-structured | invoices, purchase orders |
Custom neural models share the same labeling format and strategy as custom template models. Currently custom neural models only support a subset of the field types supported by custom template models.
Model capabilities
Important
Starting with API version 2024-02-29-preview
custom neural models add support for overlapping fields and table cell confidence.
Custom neural models currently support key-value pairs and selection marks and structured fields (tables).
Form fields | Selection marks | Tabular fields | Signature | Region labeling | Overlapping fields |
---|---|---|---|---|---|
Supported | Supported | Supported | Unsupported | Supported 1 | Supported 2 |
1 Region labels in custom neural models use the results from the Layout API for specified region. This feature is different from template models where, if no value is present, text is generated at training time.
2 Overlapping fields are supported starting with REST API version 2024-02-29-preview
. Overlapping fields have some limits. For more information, see overlapping fields.
Build mode
The Build
operation supports template and neural custom models. Previous versions of the REST API and client libraries only supported a single build mode that is now known as the template mode.
Neural models support documents that have the same information, but different page structures. Examples of these documents include United States W2 forms, which share the same information, but can vary in appearance across companies. For more information, see Custom model build mode.
Overlapping fields
With the release of API versions 2024-02-29-preview
and later, custom neural models support overlapping fields:
Overlapping fields
With the release of API versions 2024-07-31-preview and later, custom neural models will support overlapping fields:
To use the overlapping fields, your dataset needs to contain at least one sample with the expected overlap. To label an overlap, use region labeling to designate each of the spans of content (with the overlap) for each field. Labeling an overlap with field selection (highlighting a value) fails in the Studio as region labeling is the only supported labeling tool for indicating field overlaps. Overlap support includes:
- Complete overlap. The same set of tokens are labeled for two different fields.
- Partial overlap. Some tokens belong to both fields, but there are tokens that are only part of one field or the other.
Overlapping fields have some limits:
- Any token or word can only be labeled as two fields.
- overlapping fields in a table can't span table rows.
- Overlapping fields can only be recognized if at least one sample in the dataset contains overlapping labels for those fields.
To use overlapping fields, label your dataset with the overlaps and train the model with the API version 2024-02-29-preview
or later.
Tabular fields
With the release of API versions 2022-06-30-preview and later, custom neural models support tabular fields (tables) to analyze table, row, and cell data with added confidence:
- Models trained with API version 2022-06-30-preview, or later will accept tabular field labels.
- Documents analyzed with custom neural models using API version 2022-06-30-preview or later will produce tabular fields aggregated across the tables.
- The results can be found in the
analyzeResult
object'sdocuments
array that is returned following an analysis operation.
Tabular fields support cross page tables by default:
- To label a table that spans multiple pages, label each row of the table across the different pages in a single table.
- As a best practice, ensure that your dataset contains a few samples of the expected variations. For example, include samples where the entire table is on a single page and where tables span two or more pages.
Tabular fields are also useful when extracting repeating information within a document that isn't recognized as a table. For example, a repeating section of work experiences in a resume can be labeled and extracted as a tabular field.
Tabular fields provide table, row and cell confidence starting with the 2024-02-29-preview
API:
Fixed or dynamic tables add confidence support for the following elements:
- Table confidence, a measure of how accurately the entire table is recognized.
- Row confidence, a measure of recognition of an individual row.
- Cell confidence, a measure of recognition of an individual cell.
The recommended approach is to review the accuracy in a top-down manner starting with the table first, followed by the row and then the cell. See confidence and accuracy scores to learn more about table, row, and cell confidence.
Supported languages and locales
See our Language Support—custom models for a complete list of supported languages.
Supported regions
As of October 18, 2022, Document Intelligence custom neural model training will only be available in the following Azure regions until further notice:
- Australia East
- Brazil South
- Canada Central
- Central India
- Central US
- East Asia
- East US
- East US2
- France Central
- Japan East
- South Central US
- Southeast Asia
- UK South
- West Europe
- West US2
- US Gov Arizona
- US Gov Virginia
Tip
You can copy a model trained in one of the select regions listed to any other region and use it accordingly.
Use the REST API or Document Intelligence Studio to copy a model to another region.
Tip
You can copy a model trained in one of the select regions listed to any other region and use it accordingly.
Use the REST API or Document Intelligence Studio to copy a model to another region.
Tip
You can copy a model trained in one of the select regions listed to any other region and use it accordingly.
Use the REST API or Document Intelligence Studio to copy a model to another region.
Input requirements
For best results, provide one clear photo or high-quality scan per document.
Supported file formats:
Model PDF Image:
jpeg/jpg
,png
,bmp
,tiff
,heif
Microsoft Office:
Word (docx), Excel (xlsx), PowerPoint (pptx), and HTMLRead ✔ ✔ ✔ Layout ✔ ✔ ✔ (2024-02-29-preview, 2023-10-31-preview, or later) General Document ✔ ✔ Prebuilt ✔ ✔ Custom neural ✔ ✔ ✱ Microsoft Office files are currently not supported for other models or versions.
For PDF and TIFF, up to 2,000 pages can be processed (with a free tier subscription, only the first two pages are processed).
The file size for analyzing documents is 500 MB for paid (S0) tier and 4 MB for free (F0) tier.
Image dimensions must be between 50 x 50 pixels and 10,000 px x 10,000 pixels.
If your PDFs are password-locked, you must remove the lock before submission.
The minimum height of the text to be extracted is 12 pixels for a 1024 x 768 pixel image. This dimension corresponds to about
8
-point text at 150 dots per inch.For custom model training, the maximum number of pages for training data is 500 for the custom template model and 50,000 for the custom neural model.
For custom extraction model training, the total size of training data is 50 MB for template model and 1G-MB for the neural model.
For custom classification model training, the total size of training data is
1GB
with a maximum of 10,000 pages.
Best practices
Custom neural models differ from custom template models in a few different ways. The custom template or model relies on a consistent visual template to extract the labeled data. Custom neural models support structured, and semi-structured to extract fields. When you're choosing between the model types, start with a neural model, and test to determine if it supports your functional needs.
- Dealing with variations - Custom neural models can generalize across different formats of a single document type. As a best practice, create a single model for all variations of a document type. Add at least five labeled samples for each of the different variations to the training dataset.
- Field naming - When you label the data, labeling the field relevant to the value improves the accuracy of the key-value pairs extracted. For example, for a field value containing the supplier ID, consider naming the field supplier_id. Field names should be in the language of the document.
- Labeling contiguous values - Value tokens/words of one field must be either:
- In a consecutive sequence in natural reading order, without interleaving with other fields
- In a region that don't cover any other fields
- Representative data - Values in training cases should be diverse and representative. For example, if a field is named date, values for this field should be a date. Synthetic value like a random string can affect model performance.
Current Limitations
- Custom neural model doesn't recognize values split across page boundaries.
- Custom neural unsupported field types are ignored if a dataset labeled for custom template models is used to train a custom neural model.
- Custom neural models are limited to 20 build operations per month. Open a support request if you need the limit increased. For more information, see Document Intelligence service quotas and limits.
Training a model
Custom neural models are available in the v3.0 and later models.
Document Type | REST API | SDK | Label and Test Models |
---|---|---|---|
Custom document | Document Intelligence 3.1 | Document Intelligence SDK | Document Intelligence Studio |
The Build
operation to train model supports a new buildMode
property, to train a custom neural model, set the buildMode
to neural
.
https://{endpoint}/documentintelligence/documentModels:build?api-version=2024-07-31-preview
{
"modelId": "string",
"description": "string",
"buildMode": "neural",
"azureBlobSource":
{
"containerUrl": "string",
"prefix": "string"
}
}
https://{endpoint}/formrecognizer/documentModels:build?api-version=v3.1:2023-07-31
{
"modelId": "string",
"description": "string",
"buildMode": "neural",
"azureBlobSource":
{
"containerUrl": "string",
"prefix": "string"
}
}
https://{endpoint}/formrecognizer/documentModels/{modelId}:copyTo?api-version=2022-08-31
{
"modelId": "string",
"description": "string",
"buildMode": "neural",
"azureBlobSource":
{
"containerUrl": "string",
"prefix": "string"
}
}
Billing
Starting with version 2024-07-31-preview
, you can train your custom neural model for longer durations than the standard 30 minutes. Previous versions are limited to 30 minutes per training instance, with a total of 20 free training instances per month. Now with 2024-07-31-preview
, you can receive 10 hours of free model training, and train a model for as long as 10 hours.
You can choose to spend all of 10 free hours on a single model build with a large set of data, or utilize it across multiple builds by adjusting the maximum duration value for the build
operation by specifying maxTrainingHours
:
POST https://{endpoint}/documentintelligence/documentModels:build?api-version=2024-07-31-preview
{
"modelId": "string",
"description": "string",
"buildMode": "neural",
...,
"maxTrainingHours": 10
}
Important
- If you would like to train additional neural models or train models for a longer time period that exceed 10 hours, billing charges apply. For details on the billing charges, refer to the pricing page.
- You can opt in for this paid training service by setting the
maxTrainingHours
to the desired maximum number of hours. API calls with no budget but with themaxTrainingHours
set as over 10 hours will fail. - As each build takes different amount of time depending on the type and size of the training dataset, billing is calculated for the actual time spent training the neural model, with a minimum of 30 minutes per training job.
- This paid training feature enables you to train larger data sets for longer durations with flexibility in the training hours.
GET /documentModels/{myCustomModel}
{
"modelId": "myCustomModel",
"trainingHours": 0.23,
"docTypes": { ... },
...
}
Note
For Document Intelligence versions v3.1 (2023-07-31)
and v3.0 (2022-08-31)
, custom neural model's paid training is not enabled. For the two older versions, you will get a maximum of 30 minutes training duration per model. If you would like to train more than 20 model instances, you can create an Azure support ticket to increase in the training limit.
Billing
For Document Intelligence versions v3.1 (2023-07-31) and v3.0 (2022-08-31)
, you receive a maximum 30 minutes of training duration per model, and a maximum of 20 trainings for free per month. If you would like to train more than 20 model instances, you can create an Azure support ticket to increase in the training limit. For Azure support ticket, enter in the summary
field: Increase Document Intelligence custom neural training (TPS) limit
.
Important
- When increasing the training limit, note that 2 custom neural model training sessions will be considered as 1 training hour. For more details on the pricing for increasing the number of training sessions, refer to the pricing page.
- Azure support ticket for training limit increase can only apply at a resource-level, not a subscription level. You can request a training limit increase for a single Document Intelligence resource by specifying your resource ID and region in the support ticket.
If you want to train models for longer durations than 30 minutes, we support paid training with our newest version, v4.0 (2024-07-31-preview)
. Using the latest version, you can train your model for a longer duration to process larger documents. For more information about paid training, see Billing v4.0.
Billing
For Document Intelligence versions v3.1 (2023-07-31) and v3.0 (2022-08-31)
, you receive a maximum 30 minutes of training duration per model, and a maximum of 20 trainings for free per month. If you would like to train more than 20 model instances, you can create an Azure support ticket to increase in the training limit. For Azure support ticket, enter in the summary
field: Increase Document Intelligence custom neural training (TPS) limit
.
Important
- When increasing the training limit, note that 2 custom neural model training sessions will be considered as 1 training hour. For more details on the pricing for increasing the number of training sessions, refer to the pricing page.
- Azure support ticket for training limit increase can only apply at a resource-level, not a subscription level. You can request a training limit increase for a single Document Intelligence resource by specifying your resource ID and region in the support ticket.
If you want to train models for longer durations than 30 minutes, we support paid training with our newest version, v4.0 (2024-07-31)
. Using the latest version, you can train your model for a longer duration to process larger documents. For more information about paid training, see Billing v4.0.
Next steps
Learn to create and compose custom models:
Pripomienky
https://aka.ms/ContentUserFeedback.
Pripravujeme: V priebehu roka 2024 postupne zrušíme službu Problémy v službe GitHub ako mechanizmus pripomienok týkajúcich sa obsahu a nahradíme ju novým systémom pripomienok. Ďalšie informácie nájdete na stránke:Odoslať a zobraziť pripomienky pre