Domain 5 of 5 · Chapter 2 of 2

Document content extraction with Content Understanding

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing AI-103 premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • Content Understanding, the umbrella over extraction
  • Content extraction analyzers: read, layout, digitalParse
  • RAG-ready output and the analyzer as a tool
  • Standard and pro analyzer modes
  • Confidence, grounding, and review routing
  • Classifying and splitting a multi-document file
  • From prebuilt to custom: the build order
  • Content Understanding, Document Intelligence, or DIY
  • Exam-pattern recognition

Choosing the extraction route: Content Understanding, Document Intelligence, or a hand-built pipeline

RequirementContent Understanding analyzerDocument Intelligence modelHand-built LLM pipeline
Confidence and groundingOpt-in per analyzer or field; grounding locates each value in the sourceReturned by prebuilt and custom modelsNone built in; accept all, review all, or build your own signal
Building a custom extractorZero-shot from plain-language field descriptions, labels added later to raise accuracyTrained from labeled samples, as few as fiveA prompt and schema you write and maintain
Template and language variationGeneralizes across it without labelingA labeled neural model per structured document typeDepends entirely on your prompt and testing
On-premises or air-gappedNo container option todayContainers, the only offline extraction pathSelf-hosted, but with no built-in confidence
Cross-document reasoningPro mode reasons across an input set against reference dataNot offered; one document per model callOnly what you build yourself
Best fitUnstructured, multimodal, or template-varied documentsHighly structured, consistent forms where accuracy matters mostA gap no analyzer covers, when you own the accuracy work

Decision tree

Must it run on-premisesor air-gapped?YesDocument Intelligencecontainers, only offline pathNoWhat do you needfrom the document?Content or layoutRAG-ready textSpecific fieldsHow much pagestructure?prebuilt-documentSearchRAG-ready Markdown outputWhich field-extraction route?prebuilt-readplain OCR textprebuilt-layouttables, sections, marksprebuilt-digitalParseborn-digital internalsPrebuilt domain analyzere.g. prebuilt-invoiceContent Understandingpro mode, reason across docsDocument Intelligencecustom neural model, labeledContent Understandingcustom analyzer, zero-shotAir-gapped forces Document Intelligence containers; otherwise route by what you need out of the document.
Cheat sheet

Cheat sheet

  • Pro mode is the only Content Understanding mode that takes several input documents at once and reasons against reference data
  • Pro mode gives up confidence scores, grounding, and extract fields to buy multi-step reasoning
  • Standard mode is the default, tuned for cost-effective, low-latency schema extraction with optional labeling
  • Pro mode consults reference documents in lookup mode, so anything needing exhaustive recall belongs in the input set
  • contentCategories plus enableSegment splits one multi-document file into per-category page ranges and routes each segment to its own analyzer
  • Confidence and grounding are opt-in flags on the analyzer, not on by default
  • Per-field confidence exists so that only the values under your threshold go to a human reviewer
  • Grounding returns where in the source document each extracted value came from
  • A hand-built LLM extraction pipeline has no confidence signal, so every value is either trusted or reviewed
  • prebuilt-layout, prebuilt-read, and prebuilt-digitalParse are three different depths of content extraction
  • Structure analysis keeps complex tables, selection marks, barcodes, and LaTeX formulas intact
  • Prebuilt analyzer definitions can change between API versions, so production pins them with the copy operation
  • The pure content-extraction analyzers require neither a language model nor an embedding model
  • prebuilt-documentSearch is the analyzer Microsoft recommends for document ingestion into RAG
  • A custom analyzer must declare a baseAnalyzerId naming one of exactly four base analyzers
  • An analyzer can be invoked as a tool inside an agentic flow, returning a strongly typed structured representation
  • prebuilt-documentFieldSchema proposes a field schema for a document type you have not modelled yet
  • A Content Understanding custom analyzer is zero-shot from field descriptions; a Document Intelligence custom model is trained from labels
  • Document Intelligence containers are the only on-premises or air-gapped document extraction option
  • If a prebuilt covers the document type, Microsoft's guidance is to start there before building anything custom
  • Content Understanding now encompasses Document Intelligence, and existing Document Intelligence workloads need no migration

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. What is Azure AI Content Understanding?
  2. What is Azure AI Document Intelligence?
  3. Azure AI Content Understanding prebuilt analyzers
  4. Choose the right Azure AI tool for document processing
  5. Azure AI Content Understanding standard and pro modes
  6. Azure AI Content Understanding analyzer configuration and reference
  7. Install and run Docker containers for Document Intelligence
  8. Custom neural document model (Document Intelligence)