Azure Document Intelligence Custom Models: Training for Specialized Documents

When prebuilt models do not capture your document structure, custom models fill the gap. Azure Document Intelligence custom models learn from your labeled examples to extract fields specific to your business documents.

Choosing the Right Custom Model Type

Azure offers three custom model approaches: template models for fixed layouts, neural models for varied layouts, and composed models that combine multiple extractors.

Preparing Training Data

import os
import json
from azure.ai.documentintelligence import DocumentIntelligenceAdministrationClient
from azure.core.credentials import AzureKeyCredential

admin_client = DocumentIntelligenceAdministrationClient(
    endpoint=os.environ["AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT"],
    credential=AzureKeyCredential(os.environ["AZURE_DOCUMENT_INTELLIGENCE_KEY"])
)

def prepare_labeling_project(
    project_name: str,
    container_sas_url: str,
    field_schema: dict
) -> dict:
    """Set up a labeling project for custom model training."""

    project_config = {
        "projectName": project_name,
        "storageSource": {
            "kind": "AzureBlobContainer",
            "containerSasUrl": container_sas_url
        },
        "fieldSchema": field_schema
    }

    return project_config

# Define fields to extract from purchase orders
po_schema = {
    "fields": {
        "PurchaseOrderNumber": {"type": "string"},
        "OrderDate": {"type": "date"},
        "VendorName": {"type": "string"},
        "VendorAddress": {"type": "address"},
        "ShipToAddress": {"type": "address"},
        "LineItems": {
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "ItemNumber": {"type": "string"},
                    "Description": {"type": "string"},
                    "Quantity": {"type": "number"},
                    "UnitPrice": {"type": "currency"},
                    "TotalPrice": {"type": "currency"}
                }
            }
        },
        "Subtotal": {"type": "currency"},
        "Tax": {"type": "currency"},
        "Total": {"type": "currency"}
    }
}

Training the Custom Model

def train_custom_model(
    model_id: str,
    training_data_sas_url: str,
    model_description: str
) -> str:
    """Train a custom neural document model."""

    poller = admin_client.begin_build_document_model(
        build_request={
            "modelId": model_id,
            "description": model_description,
            "buildMode": "neural",  # Use neural for varied layouts
            "azureBlobSource": {
                "containerUrl": training_data_sas_url
            }
        }
    )

    model = poller.result()

    print(f"Model ID: {model.model_id}")
    print(f"Status: {model.status}")
    print(f"Created: {model.created_on}")
    print(f"Document types: {list(model.doc_types.keys())}")

    return model.model_id

def evaluate_model(model_id: str, test_document_url: str) -> dict:
    """Evaluate model performance on a test document."""

    from azure.ai.documentintelligence import DocumentIntelligenceClient

    client = DocumentIntelligenceClient(
        endpoint=os.environ["AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT"],
        credential=AzureKeyCredential(os.environ["AZURE_DOCUMENT_INTELLIGENCE_KEY"])
    )

    poller = client.begin_analyze_document(
        model_id=model_id,
        analyze_request={"urlSource": test_document_url}
    )

    result = poller.result()

    # Calculate field-level confidence
    confidence_scores = {}
    for doc in result.documents:
        for field_name, field in doc.fields.items():
            confidence_scores[field_name] = field.confidence

    return {
        "overall_confidence": sum(confidence_scores.values()) / len(confidence_scores),
        "field_confidences": confidence_scores
    }

Best Practices

Provide at least 5 labeled samples per document variation. Include edge cases like partially filled forms and poor scan quality. Neural models require more samples (15-20) but handle layout variations better than template models.

Custom models unlock document automation for your unique business documents that no prebuilt model can handle.