Skip to content

Implement an Azure AI Document Intelligence Solution

This section of the Microsoft AI-102: Designing and Implementing a Microsoft Azure AI Solution exam covers building and managing Azure AI Document Intelligence solutions. Below are study notes for each sub-topic, with links to Microsoft documentation, exam tips, and key facts


Provision a Document Intelligence Resource

📖 Docs: Create a Document Intelligence resource

Overview

  • Document Intelligence (formerly Form Recognizer) extracts structured data from documents
  • Provision in the Azure portal, CLI, or ARM template
  • Resource includes endpoint + key for API access

Key Points

  • Regional availability varies
  • Pricing based on pages processed
  • RBAC controls access

Exam Tip

Always provision Document Intelligence resource, not generic Cognitive Services


Use Prebuilt Models to Extract Data from Documents

📖 Docs: Document processing models

Overview

  • Prebuilt models handle common document types:
    • Invoices
    • Receipts
    • Business cards
    • Identity documents
    • Layout (general text extraction)

Key Points

  • Quick start without training
  • Outputs structured JSON with fields + confidence scores
  • Useful for automating business workflows

Use Case

Extracting vendor, total, and date from an invoice


Implement a Custom Document Intelligence Model

📖 Docs: Document Intelligence custom models

Overview

  • Custom models trained on user-provided documents
  • Suitable when prebuilt models don’t fit requirements

Key Points

  • Requires labeled training data
  • Labeling done via Document Intelligence Studio
  • Handles semi-structured or unstructured forms

Train, Test, and Publish a Custom Document Intelligence Model

📖 Docs: Build and train a custom extraction model

Overview

  • Training creates a custom model from labeled docs
  • Evaluation validates accuracy
  • Publishing exposes endpoint for consumption

Key Points

  • Supports iterative retraining for improvement
  • Metrics include accuracy, recall, and precision
  • Must publish before calling via API

Exam Tip

“Model trained but not available to apps” → must publish model


Create a Composed Document Intelligence Model

📖 Docs: Document Intelligence composed custom models

Overview

  • Composed model = combines multiple custom models
  • Automatically selects correct sub-model at runtime

Key Points

  • Useful when multiple document types exist
  • Reduces need to pre-sort documents
  • Example: invoices + receipts + tax forms

Use Case

A composed model that routes between invoice and receipt extractors


Quick‑fire revision sheet

  • 📌 Provision Document Intelligence resource, region + pricing matter
  • 📌 Prebuilt models = invoices, receipts, IDs, business cards, layout
  • 📌 Custom models require labeled data via Studio
  • 📌 Must train, test, publish before API consumption
  • 📌 Composed model = collection of multiple models auto-selected at runtime