Skip to content
Cheat sheets

AI-103 · Domain 5: Implement information extraction solutions · 13% of the exam

Task 5.2: Extract content from documents

Multimodal extraction that combines OCR, layout and fields, and Content Understanding analyzers that return grounded, structured or Markdown output for agents and RAG.

Study it

  • OCR, layout and field extraction

    Lesson coming

  • Content Understanding analyzers with structured and Markdown output

    Lesson coming

Sample questions

Try each one before opening the answer. Every option is explained, with the Microsoft documentation page that proves it.

Question 1 · choose 1

A Content Understanding analyzer extracts invoice totals and due dates. Before payment, reviewers must check low-confidence values and jump to the exact spot on the page where each value was found. Which analyzer setting supports this?

  1. ASet estimateFieldSourceAndConfidence to true
  2. BSet the method of every field to generate
  3. CSet returnDetails to true and drop the field schema
  4. DSet omitContent to true to shrink the response
Show the answer and why

  • ASet estimateFieldSourceAndConfidence to true

    Correct

    It returns each field's source location (page and bounding box) and a confidence score, for validation and quality assurance workflows.

  • BSet the method of every field to generate

    Incorrect

    The field method sets how a value is produced; it doesn't turn on source locations or confidence scores.

  • CSet returnDetails to true and drop the field schema

    Incorrect

    Without a field schema there are no invoice fields to locate or score.

  • DSet omitContent to true to shrink the response

    Incorrect

    Omitting content reduces output; it adds no locations or scores.

Grounded extraction means every value can be traced. Turn on source and confidence estimation (it's required for fields that use the extract method) and route low-confidence fields to human review.

Microsoft documentation

Task 5.2Report a problem with this question

Question 2 · choose 1

A knowledge assistant must ingest PDF reports full of tables, charts and diagrams into a vector index. The team wants structure kept as Markdown, charts described in words and chunks ready to embed, with no custom code. Which analyzer should it use?

  1. AThe prebuilt-read analyzer
  2. BThe prebuilt-audioSearch analyzer
  3. CA custom analyzer derived from prebuilt-image
  4. DThe prebuilt-documentSearch analyzer
Show the answer and why

  • AThe prebuilt-read analyzer

    Incorrect

    prebuilt-read gives basic OCR text without layout analysis or figure descriptions.

  • BThe prebuilt-audioSearch analyzer

    Incorrect

    This analyzer targets audio and conversation content, not documents.

  • CA custom analyzer derived from prebuilt-image

    Incorrect

    Image analyzers target pictures, not multi-page documents with tables and text.

  • DThe prebuilt-documentSearch analyzer

    Correct

    It extracts layout as Markdown, describes figures and charts, adds a summary and returns chunked output ready for vector indexing.

prebuilt-documentSearch turns documents into structured Markdown with figure descriptions and embedding-ready chunks, so they can be indexed without custom code.

Microsoft documentation

Task 5.2Report a problem with this question

Question 3 · choose 1

Suppliers upload single PDFs that each contain several invoices and receipts, plus the odd cover letter. Each invoice and receipt must be split out and sent to its own extraction analyzer. How should the analyzer be configured?

  1. AUse contentCategories without enableSegment and list two categories
  2. BSegmented contentCategories, linked analyzers and an "other" category
  3. CSet segmentPerPage on the invoice analyzer and skip classification
  4. DRun prebuilt-layout on the bundle and split it on every heading
Show the answer and why

  • AUse contentCategories without enableSegment and list two categories

    Incorrect

    Without segmentation the whole file gets one category, and with no "other" category every page is forced into invoice or receipt.

  • BSegmented contentCategories, linked analyzers and an "other" category

    Correct

    Segmentation splits the file by category, linked analyzers process each part, and an "other" category catches content such as cover letters.

  • CSet segmentPerPage on the invoice analyzer and skip classification

    Incorrect

    Page segments don't follow document boundaries and aren't routed to the receipt analyzer.

  • DRun prebuilt-layout on the bundle and split it on every heading

    Incorrect

    Headings don't reliably mark where one document ends, and layout output alone doesn't classify or route.

For mixed batches, define categories with clear descriptions, enable segmentation and point each category at an analyzer. Always add an "other" category so unmatched content isn't forced into a wrong one.

Microsoft documentation

Task 5.2Report a problem with this question

Question 4 · choose 2

You are writing the field schema for a custom Content Understanding analyzer for supplier invoices. InvoiceNumber must be returned exactly as printed, and PaymentRisk must be one of low, medium or high. Which two field definitions are correct? (Choose TWO.)

  1. AInvoiceNumber with method generate and a clear description
  2. BPaymentRisk as an enum with method classify
  3. CA Summary field with method classify and no enum values
  4. DPaymentRisk with method extract and source estimation on
  5. EInvoiceNumber with method extract and source estimation on
Show the answer and why

  • AInvoiceNumber with method generate and a clear description

    Incorrect

    generate writes values freely with AI models, which suits interpretation, not a literal identifier.

  • BPaymentRisk as an enum with method classify

    Correct

    classify picks from a predefined set of categories, which fits an enum of three values.

  • CA Summary field with method classify and no enum values

    Incorrect

    classify needs a fixed set of values; a free summary is a generate field.

  • DPaymentRisk with method extract and source estimation on

    Incorrect

    The risk level isn't printed on the invoice, so there's nothing literal to extract.

  • EInvoiceNumber with method extract and source estimation on

    Correct

    extract returns values as they appear in the content and requires source and confidence estimation for the field.

Field methods match the kind of value: extract for literal text at a location, classify for a fixed set of options, generate for values that need interpretation such as summaries.

Microsoft documentation

Task 5.2Report a problem with this question

Question 5 · choose 1

A defense contractor must extract fields from documents on a network with no connection to the cloud. Which option does the document processing guidance name for this?

  1. ADocument Intelligence containers deployed on premises
  2. BThe Content Understanding prebuilt-documentSearch analyzer
  3. CA Foundry agent with the file search tool
  4. DA Content Understanding custom analyzer called over the internet
Show the answer and why

  • ADocument Intelligence containers deployed on premises

    Correct

    For on-premises or air-gapped deployment, DI containers are the only option today.

  • BThe Content Understanding prebuilt-documentSearch analyzer

    Incorrect

    It is a cloud analyzer, unavailable without connectivity.

  • CA Foundry agent with the file search tool

    Incorrect

    Agents and their tools run in the cloud service.

  • DA Content Understanding custom analyzer called over the internet

    Incorrect

    A cloud API call isn't possible from an air-gapped network.

Deployment constraints can decide the tool before features do: when data can't leave the site, use Document Intelligence containers.

Microsoft documentation

Task 5.2Report a problem with this question

Question 6 · choose 1

A team spent weeks tuning a custom Content Understanding analyzer in its development resource. It now needs the same analyzer in a production resource in another subscription, and a backup for failover. What should it do?

  1. AExport the analyzer's results and import them as a new analyzer
  2. BPoint production apps at the development resource
  3. CUse the analyzer copy operation to copy it across resources
  4. DRecreate the analyzer by hand from screenshots of its settings
Show the answer and why

  • AExport the analyzer's results and import them as a new analyzer

    Incorrect

    Results aren't an analyzer definition.

  • BPoint production apps at the development resource

    Incorrect

    That ties production to a development resource and gives no backup.

  • CUse the analyzer copy operation to copy it across resources

    Correct

    Copy works within a resource for backups and across resources, subscriptions and tenants for failover and sharing.

  • DRecreate the analyzer by hand from screenshots of its settings

    Incorrect

    Manual rebuilding risks drift from the tuned analyzer.

Treat analyzers as assets: copy them for backup, iteration, promotion across environments and failover.

Microsoft documentation

Task 5.2Report a problem with this question

Question 7 · choose 1

An analyst wants to build a custom Content Understanding analyzer in a UI, label sample documents for in-context learning and then call it through the GA API. Which experience fits?

  1. AA Foundry agent's instructions with sample documents attached
  2. BThe Content Understanding playground in Foundry (new)
  3. CContent Understanding Studio, with its labeling tools
  4. DThe Azure AI Search Import data wizard in the portal
Show the answer and why

  • AA Foundry agent's instructions with sample documents attached

    Incorrect

    Agent instructions don't define analyzers.

  • BThe Content Understanding playground in Foundry (new)

    Incorrect

    That playground uses the preview API and doesn't create custom analyzers or support data labeling.

  • CContent Understanding Studio, with its labeling tools

    Correct

    The Studio supports the GA API, custom analyzer authoring and in-context learning with data labeling.

  • DThe Azure AI Search Import data wizard in the portal

    Incorrect

    The wizard builds search indexes, not Content Understanding analyzers.

Choose the tool by task: Content Understanding Studio for the full prebuilt catalog, custom analyzers and labeling; the Foundry playground for trying analyzers on the preview API.

Microsoft documentation

Task 5.2Report a problem with this question

Question 8 · choose 1

A financial portal wants bar and line charts found in PDF reports turned into data it can re-render with the Chart.js library. Which document analyzer setting relates to this output?

  1. AcontentCategories with a category named chart
  2. BchartFormat, returning Chart.js-compatible data
  3. CtableFormat, set to html for the charts in each report
  4. DsegmentPerPage, so that each chart gets a page of its own
Show the answer and why

  • AcontentCategories with a category named chart

    Incorrect

    Categories classify content; they don't extract chart values.

  • BchartFormat, returning Chart.js-compatible data

    Correct

    chartFormat specifies the format for extracted chart data, compatible with Chart.js, for re-rendering charts.

  • CtableFormat, set to html for the charts in each report

    Incorrect

    tableFormat controls tables, not charts.

  • DsegmentPerPage, so that each chart gets a page of its own

    Incorrect

    Page segmentation doesn't produce chart data.

Document analyzers expose output options per element type, including tables, charts and figures, so downstream code gets the format it needs.

Microsoft documentation

Task 5.2Report a problem with this question

Question 9 · choose 1

A team is about to design a custom analyzer for invoices from many vendors, field by field. A colleague points out that Content Understanding has a prebuilt for this document type. What does the guidance say?

  1. AAlways build custom analyzers to control every field
  2. BTrain a speech model on vendors' invoice terms
  3. CUse prebuilt-read and parse invoice fields with regular expressions
  4. DStart with the prebuilt analyzer, then customize if needed
Show the answer and why

  • AAlways build custom analyzers to control every field

    Incorrect

    Custom first ignores prebuilts tuned for the document type.

  • BTrain a speech model on vendors' invoice terms

    Incorrect

    Speech models don't process documents.

  • CUse prebuilt-read and parse invoice fields with regular expressions

    Incorrect

    OCR plus parsing is the custom work prebuilts remove.

  • DStart with the prebuilt analyzer, then customize if needed

    Correct

    If a prebuilt covers the document type, start there before building a custom solution; prebuilts can be customized.

Prebuilt analyzers come with field extraction for their document types; extend them with extra fields rather than starting from zero.

Microsoft documentation

Task 5.2Report a problem with this question

Practise domain 5 →Practise all domains →