Task 2.4: Implement AI solutions for information extraction by using Foundry

• AAnalyze the invoices with the prebuilt-receipt analyzer

  Incorrect

  prebuilt-receipt is tuned for sales receipts from retail and dining establishments, not supplier invoices.

• BAnalyze the invoices with the prebuilt-invoice analyzer

  Correct

  prebuilt-invoice is a domain-specific analyzer with a tuned schema for invoices, including line items, dates and totals, ready to use.

• CAnalyze the invoices with the prebuilt-documentSearch analyzer

  Incorrect

  prebuilt-documentSearch prepares documents for RAG with markdown, figure descriptions and summaries, not invoice fields.

• DAnalyze the invoices with the prebuilt-imageSearch analyzer

  Incorrect

  prebuilt-imageSearch generates descriptions of images for search. It doesn't extract invoice fields.

Content Understanding includes domain-specific prebuilt analyzers, such as prebuilt-invoice and prebuilt-receipt, that extract common fields without your own schema.

• AExtract

  Incorrect

  Extract returns values as they appear in the content. A fixed list of conditions calls for a category instead.

• BGenerate

  Incorrect

  Generate produces free-form values such as descriptions. It doesn't hold the answer to the three allowed conditions.

• CClassify

  Correct

  Classify assigns one value from a predefined set of categories, which is what a fixed list of pallet conditions needs.

• DEmbed

  Incorrect

  Embeddings are vectors used for search. Fields are produced by extract, classify or generate.

Content Understanding fields use one of three methods: extract values as they appear, classify into predefined categories, or generate values freely. A fixed set of outcomes calls for classify.

• ASwitch the analyzer to prebuilt-read to get the raw text of each form

  Incorrect

  prebuilt-read returns text without field values, confidence scores or grounding.

• BSet the locales property to auto for multilingual transcription

  Incorrect

  The locales setting controls transcription language for audio. It has nothing to do with field confidence.

• CReturn the Markdown representation of each form without fields

  Incorrect

  Markdown output suits search and RAG. Without field values and scores, nothing can be routed by reliability.

• DTurn on estimateFieldSourceAndConfidence in the analyzer

  Correct

  This setting returns a confidence score from 0 to 1 and the source region for each value, so you can route low scores to review and show reviewers where values came from.

Confidence scores and grounding let automation send high-confidence values straight through and send the rest for human review, while reviewers trace each value back to its location in the source.

• Aprebuilt-videoSearch

  Incorrect

  prebuilt-videoSearch segments videos and describes scenes. The episodes are audio only.

• Bprebuilt-audioSearch

  Correct

  prebuilt-audioSearch transcribes conversations in audio files and generates a one-paragraph summary.

• Cprebuilt-documentSearch

  Incorrect

  prebuilt-documentSearch prepares documents such as PDFs and Office files for retrieval. It doesn't transcribe audio.

• Dprebuilt-imageSearch

  Incorrect

  prebuilt-imageSearch describes images for search. It doesn't process audio.

Content Understanding's RAG analyzers include prebuilt-audioSearch, which transcribes conversations and summarizes them, alongside document, image and video counterparts.

• AIt's a long-running operation; call result() on the poller to wait for the contents

  Correct

  Content Understanding analysis is asynchronous. The SDK's poller handles the polling, and result() returns markdown and fields when the job succeeds.

• BThe call failed; retry it with recognize_once_async() on the same client

  Incorrect

  recognize_once_async() belongs to the Speech SDK's recognizer, not to Content Understanding, and a poller doesn't indicate failure.

• CThe analyzer isn't deployed yet; create a model deployment named prebuilt-invoice first

  Incorrect

  Prebuilt analyzers are called by ID. A poller is the normal response to starting an analysis, not a sign of a missing deployment.

• DThe poller is a stream of text; read its output_text property for the fields

  Incorrect

  output_text belongs to Responses API results. The poller tracks an operation and has no output_text.

Submitting content to an analyzer starts an asynchronous job. With REST you poll the Operation-Location URL until the status is Succeeded; the Python SDK wraps this as a poller whose result() waits and returns the analysis.

• AThe prebuilt-read analyzer, to extract the text of each page in the file

  Incorrect

  Reading the text of each page doesn't decide which pages belong to which document type.

• BA generate field that writes a one-paragraph summary of the whole PDF file

  Incorrect

  A summary describes the file as a whole. It doesn't split it or route its parts.

• CClassification with segmentation, with each category linked to an analyzer

  Correct

  Classify-and-segment splits a file that holds several documents, labels each part and routes it to the analyzer linked to its category.

• DThe prebuilt-documentSearch analyzer, to chunk the PDF for vector search

  Incorrect

  Chunking prepares content for retrieval in RAG. Chunks aren't classified by document type or routed for extraction.

Content Understanding classification can classify a whole file or split a file that combines several documents. Linking categories to analyzers routes each segment to the right extraction analyzer in one operation.

• AA new thumbnail image generated for each segment

  Incorrect

  Creating new images is a job for an image generation model. The analyzer describes what's in the video.

• BIdentity matches for each face against an employee photo list

  Incorrect

  Matching faces to known people is facial recognition in the Face service, which has limited access. The video analyzer doesn't do it.

• CA narration track read aloud in a neural voice

  Incorrect

  Creating narration from text is text to speech, not a Content Understanding output.

• DA transcript of the speech in WebVTT format

  Correct

  Video analysis includes the transcript in standard WebVTT format, with timestamps.

• ESegments based on scene or topic changes, each with a natural-language description

  Correct

  The analyzer segments the video automatically and describes each segment, ready for indexing.

The video analyzers return RAG-ready output: WebVTT transcripts, key frames and segment descriptions created through automatic scene segmentation, which can go straight into a search index.

• AA translated copy of the audio in a second language

  Incorrect

  Producing translated audio isn't an audio analyzer output.

• BA one-paragraph summary of each call and its topics

  Incorrect

  Analyzers can generate summaries and topics, but those describe the whole call and don't give the timing of individual words.

• CA sentiment value for each call as a whole

  Incorrect

  Sentiment can be extracted as a field, but a call-level value doesn't locate when any word was spoken.

• DWord-level timestamps in the transcript, on request

  Correct

  Audio analyzers return transcripts in WebVTT format, with sentence- and word-level timestamps available on request.

Content Understanding audio analyzers transcribe audio into WebVTT transcripts and can include sentence-level and word-level timestamps when you request them. Generated fields such as summaries, topics and sentiment describe the call as a whole.

• AOperation-Location

  Incorrect

  Operation-Location is a response header that points to the operation to poll, not a credential.

• BOcp-Apim-Subscription-Key

  Correct

  Content Understanding REST examples send the resource key in the Ocp-Apim-Subscription-Key header.

• CContent-Type

  Incorrect

  Content-Type describes the body format, such as JSON.

• DAuthorization: Bearer

  Incorrect

  A bearer token in the Authorization header is used with Microsoft Entra ID authentication, not with the resource key.

With key-based authentication, Content Understanding REST requests include the Foundry resource key in the Ocp-Apim-Subscription-Key header, alongside a Content-Type header for the JSON body.

• Aprebuilt-documentFields

  Correct

  This utility analyzer extracts key-value pairs from documents, and domain-specific analyzers use it when input doesn't match their schemas.

• Bprebuilt-contract

  Incorrect

  prebuilt-contract extracts a fixed set of fields from business contracts and agreements. It's tuned to one document category.

• Cprebuilt-documentSearch

  Incorrect

  prebuilt-documentSearch produces Markdown, figure descriptions, a summary and chunks for retrieval. It isn't built to return key-value pairs.

• Dprebuilt-procurement

  Incorrect

  prebuilt-procurement classifies procurement documents such as invoices and receipts and routes each one to its own analyzer. Other kinds of documents don't fit it.

Content Understanding's utility analyzers include prebuilt-documentFields, which extracts key-value pairs from documents and is also used by domain-specific analyzers when input doesn't match their schemas. Domain-specific analyzers such as prebuilt-contract target one document category, and prebuilt-documentSearch prepares content for RAG.

• AWrite an automated test suite against the REST API before trying it

  Incorrect

  Tests are useful later, but writing a suite before seeing any output is the slow path. The team wants a quick look first.

• BRun the AI Red Teaming Agent against the analyzer

  Incorrect

  Red teaming probes generative AI systems for safety risks; it doesn't check extracted fields.

• CRun it on sample files in the Foundry portal and review the JSON

  Correct

  The portal lets you try analyzers on examples or your own files and view the extracted fields and JSON before automating.

• DRead the analyzer's definition JSON to confirm its field schema

  Incorrect

  The definition shows which fields the analyzer is set up to extract. It doesn't show what the analyzer returns for real documents.

Using Content Understanding in the Foundry portal is a fast way to validate that an analyzer returns the expected fields and JSON before you automate the workflow in code.