Custom extractor with generative AI

Generative AI training and extraction lets you:

• Use zero-shot and few-shot technology to get a high performing model with little to no training data using the foundation model.
• Use fine-tuning to further boost accuracy as you provide more and more training data.

Generative AI training methods

The training method you choose depends on the amount of documents you have available and the amount of effort available to put into training your model. There are three ways to train a generative AI model:

Training method	Zero-shot	Few-shot	Fine-tuning
Accuracy	Medium	Medium-high	High
Effort	Low	Low	Medium
Recommended number of training documents	0	5 to 10	10 to 50+

Note: You can expect reasonably good results with 10 to 50+ training examples, and performance can be further optimized by increasing the number of examples to hundreds or thousands, depending on the complexity and variation of the documents and extraction tasks.

Custom extractor model versions

The following models are available for custom extractor. To change model versions, see Manage processor versions.

Version 1.0 and 1.3 support confidence scores, while versions 1.1 and 1.2 don't.

Note: The labeling system for Custom Extractor has been upgraded to use the latest version of the OCR model on Sept 26, 2024. If you created your processor before then, we recommend creating a new processor and relabeling the training and evaluation documents to benefit from both the improved quality with the new processor versions of Custom Extractor (v1.2 and v1.3) and the enhanced labeling system.

Model version	Description	Stable	ML processing in US/EU	Fine-tuning in US/EU	Release date
pretrained-foundation-model-v1.0-2023-08-22	Production-ready candidate specialized for document use cases powered by specialized vision models and foundation models. Note: Effective April 9, 2025, this version will no longer be accessible.	Yes	No	US only (GA)	January 9, 2024
pretrained-foundation-model-v1.1-2024-03-12	Release candidate powered by Gemini 1.0 Pro LLM and newly developed technologies, including specialized language and vision models. This and later versions include advanced OCR features such as checkbox detection. Recommended for using the increased token limits or experimenting with newer models. Note: Effective April 9, 2025, this version will no longer be accessible.	No	Yes	US, EU (Preview)	May 3, 2024
pretrained-foundation-model-v1.2-2024-05-10	Production-ready candidate powered by Gemini 1.5 Pro LLM and newly developed technologies, including specialized language and vision models. Recommended for best quality, using the largest supported token limits, employing the best quality in identifying entities, or experimenting with newer models.	Yes	Yes	US, EU (Preview)	May 27, 2024
pretrained-foundation-model-v1.3-2024-08-31	Production-ready candidate powered by the Gemini 1.5 Flash LLM. Recommended for those who want the lowest latency.	Yes	Yes	US, EU (Preview)	Aug 23, 2024

Note: For processor versions v1.0 and v1.1, you will need to migrate to a newer version to avoid any service disruptions, such as pretrained-foundation-model-v1.2-2024-05-10 and pretrained-foundation-model-v1.3-2024-08-31 for improved quality from the latest proprietary vision models and foundation models.

To change the processor version in your project, review Managing processor versions.

Before getting started

If not done so already, enable billing and the Document AI APIs.

Build and evaluate a generative AI model

1. Create a processor and define fields you want to extract following best practices, which is important because it impacts extraction quality.

   • Go to Workbench > Custom extractor > Create processor > Assign a name.

   

   • Go to Get started > Create new field.

   

   Note: The field names with the foundation model can greatly affect model accuracy and performance. A descriptive name is recommended.

2. Import documents

   • Import documents with auto-labeling and assign documents to the training and test set.
   • For zero-shot, only the schema is required. To evaluate the accuracy of the model, only a testing set is needed.
   • For few-shot, we recommend five training documents.
   • The number of needed testing documents depend on the use case. Generally, more testing documents is better.
   • Confirm or edit the labels in the document.

3. Train model:

   • Select the Build, then Create new version.
   • Enter a name and select Create.

   

Note: Each training takes time to complete.

1. Evaluation:

   • Go to Evaluate & test, select the version you just trained, then select View full evaluation.

   

   • You now see metrics such as f1, precision, and recall for the entire document and each field.
   • Decide if performance meets your production goals. If it does not then reevaluate training and testing sets.

2. Set a new version as default:

   • Navigate to Manage versions.
   • Select to expand the options, and then select Set as default.
   Note: Ensure the Status of the version is deployed. If it is not, select the deploy version first. You can now set the new version as the default.

   

   Your model is now deployed. Documents sent to this processor use your custom version. You can evaluate the model's performance to check if it requires further training.

Evaluation reference

The evaluation engine can do both exact match or fuzzy matching. For an exact match, the extracted value must exactly match the ground truth or is counted as a miss.

Fuzzy matching extractions that had slight differences such as capitalization differences still count as a match. This can be changed at the Evaluation screen.

Fine-tuning

With fine-tuning, you use hundreds or thousands of documents for your training.

1. Create a processor and define fields you want to extract following best practices, which is important because it impacts extraction quality.

   Note: The field names with the foundation model can greatly affect model accuracy and performance. Be sure to give a descriptive name.

2. Import documents with auto-labeling, and assign documents to the training and test set.

3. Confirm or edit the labels in the document.

4. Train model.

   • Select the Build tab and select Create New Version.

   • Try out the default training parameters or values provided. If the results are unsatisfactory, experiment with these advanced options:

   • Training steps (between 100 and 400): Controls how often the weights are optimized on a batch of data during the tuning.

     • Too low indicates a risk that the training ends before convergence (under-fitting).
     • Too high means the model might see the same batch of data multiple times during the training, which can lead to overfitting.
     • Fewer steps leads to faster training time. Higher counts can help for documents with little template variation (and lower ones for those with more variation).

   • Learning rate multiplier (between 0.1 and 10): Controls how quickly the model parameters are optimized on the training data. It roughly corresponds to the size of each training step.

     • Low rates mean small changes in the model weights at each training step. If too low, the model might not converge to a stable solution.
     • High rates indicate large changes, and too high can mean the model steps over the optimal solution and converges instead to a suboptimal solution.
     • Training time is not affected by the choice of learning rate.

   • Give a name, select the required base processor version, and select Create.

   

5. Evaluation: Go to Evaluate & test, then select the version you just trained and select View full evaluation.

   

   • You now see metrics such as f1, precision, and recall for the entire document and each field.
   • Decide if performance meets your production goals. If not, then further training documents might be required.

6. Set a new version as the default:

   • Navigate to Manage versions.
   • Select to expand the options, and select Set as default.

   

   Note: Ensure the status of the version is deployed. If it's not, then instead select Deploy version. You can now set the new version as the default when its status is changed to deployed.

   Your model is now deployed and documents sent to this processor now use your custom version. You want to evaluate the model's performance to check if it requires further training.

Auto-labeling with the foundation model

The foundation model can accurately extract fields for a variety of document types, but you can also provide additional training data to improve the accuracy of the model for specific document structures.

Document AI uses the label names you define and previous annotations to make it quicker and easier to label documents at scale with auto-labeling.

1. When you've created a custom processor, go to the Get Started tab.
2. Select Create New Field.

3. Give the label a descriptive, distinct name. This will improve the foundation model's accuracy and performance.

   

4. For extraction accuracy and performance add a description (such as added context, insights, and prior knowledge for each entity) for the kinds of entities it should pick up.

Note: Regardless of content structure or variations, rules promote reliable and consistent extractions. To guide the model and minimize confusion in the document, include property descriptions with location information and text patterns of the property values.

Note: It's possible to update the description. However, the change won't be propagated to custom processor version. A snapshot of the description is in the custom processor version.

1. Navigate to the Build tab, then select Import Documents.

   

2. Select the path of the documents and which set the documents should be imported into. Check the auto-labeling option and select the foundation model.

3. In the Build tab, select Manage Dataset.

4. When you see your imported documents, select one of them.

   

The predictions from the model are now shown highlighted in purple.

1. Review each label predicted by the model, and ensure it's correct.

2. If there are missing fields, add those as well.

   Note: It's important that all fields are as accurate as possible, or model performance will suffer. More details on labeling.

   

3. When the document has been reviewed, select Mark as Labeled. The document is now ready to be used by the model.

4. Make sure the document is in either the testing or training set.

Three-level nesting

Custom Extractor now provides three levels of nesting. This feature provides better extraction for complex tables.

Note: Three levels of nesting is only supported if using generative AI ("Foundation") model types within the Custom Extractor. Extract, classify, and splitting models only support two levels of nesting.

You can determine the model type using the following API calls:

• processors.get
• processorVersions.get

The response of these is a ProcessorVersion, which contains the modelType field in v1beta3 preview.

Procedure and example

We are using this sample:

1. Select Get Started, and then create a field:

   • Create the top level.
   • In this sample, the officer_appointments is used.
   • Select This is a parent label.
   • Select Occurrence: Optional multiple.

   

2. Select Add child field. The second level label can now be created:

   • For this level label, create officer.
   • Select This is a parent label.
   • Select Occurrence: Optional multiple.

   

3. Select Add child field from second level officer. Create child labels for the third level of nesting.

4. When your schema is set, you can get predictions from documents with three levels of nesting using auto-labeling.

   

Dataset configuration

A document dataset is required to train, up-train, or evaluate a processor version. Document AI processors learn from examples, just like humans. Dataset fuels processor stability in terms of performance.

Train dataset

To improve the model and its accuracy, train a dataset on your documents. The model is made up of documents with ground-truth.

• For fine-tuning, you need a minimum of 10 documents to train a new model with version pretrained-foundation-model-v1.0-2023-08-22, 16 for version pretrained-foundation-model-v1.1-2024-03-12, and 1 document for pretrained-foundation-model-v1.2-2024-05-10 and pretrained-foundation-model-v1.3-2024-08-31.
• For a few-shot, five documents is recommended.
• For zero-shot, only a schema is required.

Ground-truth is the correctly labeled data, as determined by humans.

Test dataset

The test dataset is what the model uses to generate an F1 score (accuracy). It is made up of documents with ground-truth. To see how often the model is right, the ground truth is used to compare the model's predictions (extracted fields from the model) with the correct answers. The test dataset should have at least 10 documents for fine-tuning version pretrained-foundation-model-v1.0-2023-08-22, 16 for version pretrained-foundation-model-v1.1-2024-03-12, and 1 document for pretrained-foundation-model-v1.2-2024-05-10 and pretrained-foundation-model-v1.3-2024-08-31..

Custom extractor with property descriptions

With property descriptions, you can train a model by describing what the labeled fields are like. You can provide additional context and insights for each entity. This allows the model to train by matching fields that fit the description you provide and improve extraction accuracy. Property descriptions can be specified for both parent and child entities.

Good examples of property descriptions include location information and text patterns of the property values, which help disambiguate potential sources of confusion in the document. Clear and precise property descriptions guide the model with rules that ensure more reliable and consistent extractions, regardless of the specific document structure or content variations.

Update document schema for a processor

For how to set the property descriptions, refer to Update document schema.

Send a processing request with property descriptions

If the document schema already has descriptions set, you can send a process request with the instructions at Send a processing request.

Fine-tune a processor with property descriptions

Before using any of the request data, make the following replacements:

• LOCATION: your processor's location, for example:
  • us - United States
  • eu - European Union
• PROJECT_ID: Your Google Cloud project ID.
• PROCESSOR_ID: the ID of your custom processor.
• DISPLAY_NAME: Display name for the processor.
• PRETRAINED_PROCESSOR_VERSION: the processor version identifier. Refer to Select a processor version for more information. For example:
  • pretrained-TYPE-vX.X-YYYY-MM-DD
  • stable
  • rc
• TRAIN_STEPS: Training steps for model fine-tuning.
• LEARN_RATE_MULTIPLIER: Learning rate multiplier for model finetuning.
• DOCUMENT_SCHEMA: Schema for the processor. Refer to DocumentSchema representation.

HTTP method and URL:

POST https://LOCATION-documentai.googleapis.com/v1beta3/projects/PROJECT_ID/locations/LOCATION/processors/PROCESSOR_ID/processorVersions/PROCESSOR_VERSION:process

Request JSON body:

{
  "rawDocument": {
      "parent": "projects/PROJECT_ID/locations/LOCATION/processors/PROCESSOR_ID",
         "processor_version": {
             "name": "projects/PROJECT_ID/locations/LOCATION/processors/PROCESSOR_ID/processorVersions/DISPLAY_NAME",
             "display_name": "DISPLAY_NAME",
             "model_type": "MODEL_TYPE_GENERATIVE",
         },
         "base_processor_version": "projects/PROJECT_ID/locations/us/processors/PROCESSOR_ID/processorVersions/PRETRAINED_PROCESSOR_VERSION",
         "foundation_model_tuning_options": {
             "train_steps": TRAIN_STEPS,
             "learning_rate_multiplier": LEARN_RATE_MULTIPLIER,
         }
         "document_schema": DOCUMENT_SCHEMA
      }
}

To send your request, choose one of these options:

curl

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login , or by using Cloud Shell, which automatically logs you into the gcloud CLI . You can check the currently active account by running gcloud auth list.

Save the request body in a file named request.json, and execute the following command:

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-documentai.googleapis.com/v1beta3/projects/PROJECT_ID/locations/LOCATION/processors/PROCESSOR_ID/processorVersions/PROCESSOR_VERSION:process"

PowerShell

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login . You can check the currently active account by running gcloud auth list.

Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-documentai.googleapis.com/v1beta3/projects/PROJECT_ID/locations/LOCATION/processors/PROCESSOR_ID/processorVersions/PROCESSOR_VERSION:process" | Select-Object -Expand Content

What's next

Uptrain a specialized processor