Deploy Document Intelligence

When to Use

• Deploy Azure Document Intelligence resources and custom models
• Configure prebuilt models (invoices, receipts, ID documents, tax forms)
• Train and deploy custom extraction models
• Set up document processing pipelines with validation
• Configure PII detection and redaction on extracted content

Prerequisites

1. Azure CLI authenticated: az account show
2. Bicep CLI: az bicep version
3. Azure Document Intelligence resource (S0 tier for production)
4. Training documents for custom models (minimum 5 labeled samples)
5. Azure Storage account for document uploads

Step 1: Validate Infrastructure

az bicep lint -f infra/main.bicep
az bicep build -f infra/main.bicep

Verify resources:

• Azure Document Intelligence (S0 for production, F0 for dev)
• Azure Storage (document uploads, processed results)
• Azure OpenAI (post-extraction enrichment, if needed)
• Azure Key Vault (API keys, connection strings)
• Azure Monitor (processing metrics, error tracking)

Step 2: Deploy Azure Resources

az deployment group create \
  --resource-group $RESOURCE_GROUP \
  --template-file infra/main.bicep \
  --parameters infra/parameters.json

Step 3: Configure Prebuilt Models

Select the appropriate prebuilt model per document type:

Step 4: Train Custom Models (If Needed)

# Upload training documents
az storage blob upload-batch --source training-docs/ --destination training --account-name $STORAGE

# Start training via REST API
python scripts/train_custom_model.py --training-data $BLOB_URL --model-id custom-contract-v1

Requirements for custom models:

• Minimum 5 labeled documents (recommended 15-20)
• Consistent document layout across samples
• Clear, high-resolution scans (300 DPI minimum)
• Label all target fields in Document Intelligence Studio

Step 5: Set Up Processing Pipeline

# Document processing flow:
# 1. Upload → Blob Storage trigger
# 2. Document Intelligence analysis (prebuilt or custom model)
# 3. Field extraction with confidence scores
# 4. Validation rules check
# 5. PII detection and redaction
# 6. Store structured results
# 7. Flag low-confidence fields for human review

Step 6: Configure Validation Rules

• Field-level confidence threshold: reject < 0.80
• Cross-field validation (e.g., line item totals = invoice total)
• Required field check (reject documents missing mandatory fields)
• Format validation (dates, currency, phone numbers)

Step 7: Configure PII Handling

• Enable PII detection on all extracted text
• Redact SSN, credit card, bank account numbers
• Mask personal addresses and phone numbers in logs
• Store original only in encrypted, access-controlled storage

Step 8: Smoke Test

python scripts/test_extraction.py --document samples/invoice.pdf --model prebuilt-invoice
python scripts/test_extraction.py --document samples/receipt.jpg --model prebuilt-receipt
python scripts/test_extraction.py --document samples/contract.pdf --model custom-contract-v1

Post-Deployment Verification

• [ ] Prebuilt models returning expected fields
• [ ] Custom model accuracy ≥ 90% on validation set
• [ ] Processing time < 10 seconds per page
• [ ] PII redaction active and verified
• [ ] Low-confidence fields flagged for human review
• [ ] Blob trigger pipeline functional
• [ ] Error logging and retry logic working

Troubleshooting