Information extraction solutions · 10–15%

Extract content from documents

~2 min read

Azure Document Intelligence is a service that combines OCR with deep-learning models to pull text, key-value pairs, tables, and typed fields out of documents — returned as structured JSON with bounding boxes.

Azure Document Intelligence — model map

Category	Models	Extracts
Document analysis	read	Text (print + handwriting), language detect; Office formats; `pages` param
	layout	read + tables (cells w/ row/col index), selection marks (checkboxes), structure, optional key-value pairs
Prebuilt	invoice, receipt, ID document, W-2/1099/1040, bank statement, check, pay stub, contract, health insurance card, mortgage (1003/1004/1008)	Fixed field sets per type
Custom	template, neural, classifier, composed	Your forms

Input limits: JPEG/PNG/BMP/PDF/TIFF · ≤500 MB (free 4 MB) · 50×50 – 10,000×10,000 px · PDF ≤ A3, not password-protected.

Add-ons (some premium): high-res, formulas, fonts, barcodes, searchable PDF, query fields (natural-language field ask), key-value pairs.

Custom models — template vs neural

	Template	Neural
Layout	Fixed/consistent	Varies (semi/unstructured)
Training	Minutes, cheap	Longer, costlier
Languages	100+	Fewer
Extras	KV, selection marks, tables, signatures	Overlapping fields, signature, cell-level confidence

Training data: 5–6+ samples in Blob + SAS URL; ocr.json per form, one fields.json, labels.json per form (Studio generates these; needs CORS on storage).
A composed model bundles several custom models; it classifies the incoming document and routes it to the right component model.
Custom classifier = identify document type before extraction.
Analyze: begin_analyze_document_from_url(model_id, url) — async poller.

If the question says “fixed-layout questionnaire, cheap and fast”, the answer is template. “Invoices from many vendors, layouts vary” means neural (but check prebuilt-invoice first!). “Multiple form types, one endpoint” means composed. Always check prebuilt before training custom.

Content Understanding for agents & RAG

Content Understanding (defined in the Vision area) also serves extraction: its analyzers turn messy files into clean text agents can reason over.

Multimodal pipelines combining OCR + layout analysis + field extraction in one analyzer.
Choose the output by destination: Markdown when you want clean grounded text for RAG/search ingestion, JSON when you feed automation or agents.
Analyzers produce structured or markdown outputs for downstream reasoning; confidence + grounding minimize human review.

Document Intelligence vs Content Understanding

	Document Intelligence	Content Understanding
Modalities	Documents/images	Documents, images, audio, video
Strength	Mature prebuilts (invoice, tax, ID), custom training	Generative schema (extract/classify/generate), markdown for RAG
Pick when	Classic forms processing	Multimodal or RAG-prep extraction