Three model categories
Azure Document Intelligence uses OCR and deep learning to extract structured data — text, key-value pairs, selection marks, tables — returning JSON that preserves structure via bounding-box coordinates. It reaches accuracy without building models from scratch, leaning on models pre-trained on thousands of forms.
| Category | What it offers |
|---|---|
| Document analysis | Read (text + language detection) and Layout (adds tables and structure). |
| Prebuilt | Ready-made extractors for common types (invoices, receipts, tax forms, IDs) — no training. |
| Custom | Template, neural, composed models, and classifiers for your own document types. |
Access via REST, SDKs (Python, C#, Java, JS), the Studio, or the Foundry portal. Input: JPEG/PNG/BMP/PDF/TIFF, under 500 MB (standard tier), 50×50 to 10,000×10,000 px, not password-protected.
Explore visually, extract with prebuilt models
The Document Intelligence Studio (documentintelligence.ai.azure.com) is a visual tool to test models, review fields and confidence scores, and build custom models — auto-generating the required JSON files (ocr.json, labels.json, fields.json). Add-ons include high-resolution extraction, formula/font/barcode extraction, searchable PDF, and natural-language query fields (some are premium).
Key concept · Read, Layout, and the prebuilt catalog
The Read model extracts printed/handwritten text and detects language — best for documents with no fixed structure. The Layout model adds selection marks (checkboxes), complex tables (merged cells), structure, and optional key-value pairs. On top sit prebuilt models for financial/legal (invoice, receipt, bank statement, check, pay stub, contract), US tax (W-2, 1098/1099/1040), US mortgage (1003/1004/1008, closing disclosure), and identity (ID document, health insurance card, marriage certificate). Always check for an existing prebuilt model before building custom.
Template, neural, classifiers, composed
When no prebuilt fits, train a custom model on labeled samples (5–6+ recommended, stored in Blob Storage with a SAS URL, built via the Build model API).
| Factor | Template | Neural |
|---|---|---|
| Best for | Structured forms, consistent layout | Semi/unstructured, varying layouts |
| Training time | Minutes | Longer |
| Cost | Lower | Higher |
| Accuracy | High for fixed layouts | Higher overall, handles variation |
| Languages | 100+ | Fewer |
Key concept · Classifiers and composed models
Custom classifiers identify a document's type before extraction, routing it to the right model. A composed model bundles several custom models into one — Document Intelligence classifies the incoming document and routes it to the best component, ideal for multi-form scenarios. Start with template models for consistent layouts; move to neural if accuracy is short or layouts vary.
Chapter 29 quiz
Fifteen questions on the service, Studio/prebuilt, and custom models. Answer first, then expand Show answer.
Part A · Multiple choice
What does Document Intelligence preserve using bounding-box coordinates?
- A The document's structure and object positions
- B The file's compression
- C The user's identity
- D The billing tier
Show answer
A. Returned in JSON with coordinate tracking.
Which document-analysis model extracts text and detects language only?
- A Read
- B Layout
- C Invoice
- D Neural
Show answer
A — Read. Layout adds tables and structure.
Which model adds selection marks and complex table extraction?
- A Layout
- B Read
- C Receipt
- D Classifier
Show answer
A — Layout. It absorbed the old general-document model's features.
Which is a prebuilt model type?
- A Invoice
- B A model you must train yourself
- C Sora-2
- D FLUX
Show answer
A — Invoice. Along with receipt, W-2, ID document, and more.
Which custom model type is best for consistent, fixed layouts and trains in minutes?
- A Custom template
- B Custom neural
- C Composed
- D Read
Show answer
A — custom template. Neural handles varying layouts with higher accuracy.
What does a custom classifier do?
- A Identify a document's type before invoking extraction
- B Translate the document
- C Compress the PDF
- D Generate an image
Show answer
A. Routing documents to the right extraction model.
What does a composed model do?
- A Bundles several custom models and auto-routes documents
- B Merges two PDFs
- C Fine-tunes GPT-4.1
- D Nothing
Show answer
A. Classifies the input and routes to the best component model.
Part B · True or false
Document Intelligence can extract handwritten as well as printed text.
Show answer
True. The Read model classifies text as handwritten or printed.
Input documents can be password-protected.
Show answer
False. Password-protected files are not supported.
The Studio auto-generates ocr.json, labels.json, and fields.json for custom models.
Show answer
True. It automates the custom-model pipeline.
Custom template models support more languages than neural models.
Show answer
True. 100+ vs. fewer for neural.
You should build a custom model before checking whether a prebuilt one exists.
Show answer
False. Always check for an existing prebuilt model first.
Part C · Short answer
Name the three model categories.
Show answer
Document analysis (read/layout), prebuilt models, and custom models.
Name four things Document Intelligence can extract from a document.
Show answer
Any four: text, key-value pairs, selection marks, tables, fields, bounding boxes.
Roughly how many labeled samples are recommended to train a custom model?
Show answer
At least 5–6 — larger, more varied datasets improve accuracy.