Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 3

Question 20

A product team treats every extracted phrase as a confirmed brand or person and sends an alert. What correction is required?

  1. Apply a confidence-aware review policy instead of relying on the winning sentiment label alone
  2. Capture entity spans against the source text before applying length-changing redactions
  3. Use entity extraction to verify named-entity categories instead of interpreting key phrases as entities
  4. Use an approved terminology customization or supported custom translation workflow

Explanation

Correct Answer

C:

A salient phrase need not identify an entity; key-phrase output is not typed person or organization recognition. In the multilingual customer-support portal, confirm the fix by retesting the affected behavior after using entity extraction to verify named-entity categories instead of interpreting key phrases as entities.

Incorrect Answers

A:

Sentiment confidence reflects model uncertainty; consistent predicted labels can require different actions when confidence diverges in its own workflow, but that is a different correction from the present requirement to use entity extraction to verify named-entity categories instead of interpreting key phrases as entities. The reported issue would remain.

B:

That change addresses another failure mode: text replacement alters character positions, so spans extracted after redaction do not necessarily index the unmodified original content; it does not address this case, for which the relevant intervention is to use entity extraction to verify named-entity categories instead of interpreting key phrases as entities.

D:

Although useful elsewhere in the multilingual customer-support portal, domain terminology requires deliberate control of translated terms; generic translation alone may vary proper names across documents; the present case is different and calls for the team to use entity extraction to verify named-entity categories instead of interpreting key phrases as entities.

Question 21

The media analytics pipeline keeps processing uploaded videos but analysts see no results. The submitted job is still pending. What is the correct operational response?

  1. Route the multipage document to Document Intelligence rather than still-image captioning
  2. Add an explicit numeric claim-amount field to the analyzer extraction schema
  3. Poll or wait for the asynchronous video indexing job to complete before requesting final insights
  4. Train or adapt a supported Custom Speech model using relevant domain audio and text

Explanation

Correct Answer

C:

Video indexing is asynchronous for many workflows; a pending job is not an empty transcript and should not be treated as finished. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after polling or wait for the asynchronous video indexing job to complete before requesting final insights.

Incorrect Answers

A:

That change addresses another failure mode: document Intelligence has document and table extraction workflows; a single-image caption request is not a multipage parser; it does not address this case, for which the relevant intervention is to poll or wait for the asynchronous video indexing job to complete before requesting final insights.

B:

Although useful elsewhere in the claims intake with mixed media, structured field definitions constrain the expected output contract more directly than asking for an unconstrained narrative; the present case is different and calls for the team to poll or wait for the asynchronous video indexing job to complete before requesting final insights.

D:

Domain-adapted speech recognition can improve specialized vocabulary when supported training data is representative in its own workflow, but that is a different correction from the present requirement to poll or wait for the asynchronous video indexing job to complete before requesting final insights. The reported issue would remain.

Question 22

A team responsible for a claims intake with mixed media needs to describe visible objects and image content in uploaded photographs; the acceptance check is that the output describes objects present in the photographs. Which capability should be selected?

  1. Use Azure Content Understanding
  2. Use Azure AI Video Indexer
  3. Use Azure Vision Image Analysis
  4. Use Azure Speech in Foundry Tools

Explanation

Correct Answer

C:

Azure Vision Image Analysis supplies visual tags, captions, and supported detection results from image inputs. For this claims intake with mixed media, the intended evidence is that the output describes objects present in the photographs. It does not transcribe a telephone call or extract invoice line items into a domain-specific schema.

Incorrect Answers

A:

Content Understanding analyzers specify the input modalities and extraction schema for multimodal analysis; it may be used to produce structured insights from a combination of documents, audio, video, and images in a claims intake with mixed media, but the current requirement is instead to describe visible objects and image content in uploaded photographs. That separate capability does not prove that the output describes objects present in the photographs; It requires a defined analyzer and is broader than simple single-purpose OCR, so selecting that alternative would not meet the stated acceptance check.

B:

Video Indexer processes media to derive insights such as speech transcripts, keywords, and time-coded observations; it may be used to obtain searchable transcripts and timed insights from recorded videos in a claims intake with mixed media, but the current requirement is instead to describe visible objects and image content in uploaded photographs. That separate capability does not prove that the output describes objects present in the photographs; It is not the appropriate primary tool for a single standalone text sentence or a PDF invoice, so selecting that alternative would not meet the stated acceptance check.

D:

Speech recognition converts audio signals to transcribed words through supported real-time or batch interfaces; it may be used to turn callers' recorded speech into text for a contact-center app in a claims intake with mixed media, but the current requirement is instead to describe visible objects and image content in uploaded photographs. That separate capability does not prove that the output describes objects present in the photographs; It does not identify objects in still photos or parse expense forms, so selecting that alternative would not meet the stated acceptance check.

Question 23

A customer sends a repeat claim scan after a parser upgrade, but the stored field values differ without explanation. What control should be added?

  1. Index the transcript and its speech-derived timed segments from the video
  2. Enable a supported Read or OCR feature and consume its text spans
  3. Record and compare the extraction model or API version used for each processed document
  4. Inspect returned evidence or source grounding for that extracted date when available

Explanation

Correct Answer

C:

Document extraction output can change across model releases; versioned provenance supports meaningful regression investigations. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after recording and compare the extraction model or API version used for each processed document.

Incorrect Answers

A:

Although useful elsewhere in the claims intake with mixed media, speech transcripts support spoken-text search, while image tags describe visible scenes and cannot substitute for dialogue; the present case is different and calls for the team to record and compare the extraction model or API version used for each processed document.

B:

That change addresses another failure mode: oCR returns recognized text content and location information where supported, while generic captions are not exact text transcripts; it does not address this case, for which the relevant intervention is to record and compare the extraction model or API version used for each processed document.

D:

Source evidence permits comparison of an inferred field with the underlying document or media segment instead of trusting an isolated JSON value in its own workflow, but that is a different correction from the present requirement to record and compare the extraction model or API version used for each processed document. The reported issue would remain.

Question 24

Compliance reviewers want to jump from a detected spoken keyword to its position in a two-hour meeting recording. What information must the app preserve?

  1. Keep the returned field content and document spans or bounding regions
  2. Create modality-appropriate Content Understanding analyzers with a common field contract
  3. Use Video Indexer insights with their associated timestamps
  4. Stream microphone audio to real-time speech recognition

Explanation

Correct Answer

C:

Time-coded insights link recognized content to moments in the video, enabling direct navigation instead of a static unpositioned summary. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after using Video Indexer insights with their associated timestamps.

Incorrect Answers

A:

Document Intelligence field results can include provenance needed to inspect where the extracted value originated in its own workflow, but that is a different correction from the present requirement to use Video Indexer insights with their associated timestamps. The reported issue would remain.

B:

Although useful elsewhere in the claims intake with mixed media, a shared output schema allows downstream processing to consume related fields across supported content types despite different extractors; the present case is different and calls for the team to use Video Indexer insights with their associated timestamps.

D:

That change addresses another failure mode: real-time recognition supports interim and final transcription events during an active audio session; it does not address this case, for which the relevant intervention is to use Video Indexer insights with their associated timestamps.

Question 25

A team sends a 30-page searchable PDF to an endpoint intended for individual images and expects form tables. Which integration correction fits the input?

  1. Route the multipage document to Document Intelligence rather than still-image captioning
  2. Add an explicit numeric claim-amount field to the analyzer extraction schema
  3. Preserve document IDs, source versions, extraction model identifiers, and relevant response metadata
  4. Train or adapt a supported Custom Speech model using relevant domain audio and text

Explanation

Correct Answer

A:

Document Intelligence has document and table extraction workflows; a single-image caption request is not a multipage parser. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after routing the multipage document to Document Intelligence rather than still-image captioning.

Incorrect Answers

B:

That change addresses another failure mode: structured field definitions constrain the expected output contract more directly than asking for an unconstrained narrative; it does not address this case, for which the relevant intervention is to route the multipage document to Document Intelligence rather than still-image captioning.

C:

Model and source provenance allow analysts to compare historic results to the original document; choosing a different prebuilt model is not incident traceability in its own workflow, but that is a different correction from the present requirement to route the multipage document to Document Intelligence rather than still-image captioning. The reported issue would remain.

D:

Although useful elsewhere in the claims intake with mixed media, domain-adapted speech recognition can improve specialized vocabulary when supported training data is representative; the present case is different and calls for the team to route the multipage document to Document Intelligence rather than still-image captioning.

Question 26

Scanned purchase documents are recognized correctly, but the team cannot reproduce extraction failures from last month. What operational data should be kept?

  1. Train or adapt a supported Custom Speech model using relevant domain audio and text
  2. Poll or wait for the asynchronous video indexing job to complete before requesting final insights
  3. Add an explicit numeric claim-amount field to the analyzer extraction schema
  4. Preserve document IDs, source versions, extraction model identifiers, and relevant response metadata

Explanation

Correct Answer

D:

Model and source provenance allow analysts to compare historic results to the original document; choosing a different prebuilt model is not incident traceability. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after preserving document IDs, source versions, extraction model identifiers, and relevant response metadata.

Incorrect Answers

A:

Domain-adapted speech recognition can improve specialized vocabulary when supported training data is representative in its own workflow, but that is a different correction from the present requirement to preserve document IDs, source versions, extraction model identifiers, and relevant response metadata. The reported issue would remain.

B:

Although useful elsewhere in the claims intake with mixed media, video indexing is asynchronous for many workflows; a pending job is not an empty transcript and should not be treated as finished; the present case is different and calls for the team to preserve document IDs, source versions, extraction model identifiers, and relevant response metadata.

C:

That change addresses another failure mode: structured field definitions constrain the expected output contract more directly than asking for an unconstrained narrative; it does not address this case, for which the relevant intervention is to preserve document IDs, source versions, extraction model identifiers, and relevant response metadata.

Question 27

A live call-center screen must display words as the caller speaks, not after the call finishes. Which processing mode should the client use?

  1. Create modality-appropriate Content Understanding analyzers with a common field contract
  2. Stream microphone audio to real-time speech recognition
  3. Validate the image encoding, size, and supported input format against the selected API requirements
  4. Keep the returned field content and document spans or bounding regions

Explanation

Correct Answer

B:

Real-time recognition supports interim and final transcription events during an active audio session. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after streaming microphone audio to real-time speech recognition.

Incorrect Answers

A:

Although useful elsewhere in the claims intake with mixed media, a shared output schema allows downstream processing to consume related fields across supported content types despite different extractors; the present case is different and calls for the team to stream microphone audio to real-time speech recognition.

C:

That change addresses another failure mode: image Analysis accepts bounded input types and sizes; a valid key or region does not make every file encodable as a supported image; it does not address this case, for which the relevant intervention is to stream microphone audio to real-time speech recognition.

D:

Document Intelligence field results can include provenance needed to inspect where the extracted value originated in its own workflow, but that is a different correction from the present requirement to stream microphone audio to real-time speech recognition. The reported issue would remain.

Question 28

A release for a claims intake with mixed media is blocked until the application can turn callers' recorded speech into text for a contact-center app. The tester will verify that the deliverable is timestamped or plain speech transcription. What should be implemented?

  1. Use Azure Content Understanding
  2. Use Azure AI Video Indexer
  3. Use Azure Speech in Foundry Tools
  4. Use Azure Vision Image Analysis
  5. Use Azure Document Intelligence

Explanation

Correct Answer

C:

Speech recognition converts audio signals to transcribed words through supported real-time or batch interfaces. For this claims intake with mixed media, the intended evidence is that the deliverable is timestamped or plain speech transcription. It does not identify objects in still photos or parse expense forms.

Incorrect Answers

A:

Content Understanding analyzers specify the input modalities and extraction schema for multimodal analysis; it may be used to produce structured insights from a combination of documents, audio, video, and images in a claims intake with mixed media, but the current requirement is instead to turn callers' recorded speech into text for a contact-center app. That separate capability does not prove that the deliverable is timestamped or plain speech transcription; It requires a defined analyzer and is broader than simple single-purpose OCR, so selecting that alternative would not meet the stated acceptance check.

B:

Video Indexer processes media to derive insights such as speech transcripts, keywords, and time-coded observations; it may be used to obtain searchable transcripts and timed insights from recorded videos in a claims intake with mixed media, but the current requirement is instead to turn callers' recorded speech into text for a contact-center app. That separate capability does not prove that the deliverable is timestamped or plain speech transcription; It is not the appropriate primary tool for a single standalone text sentence or a PDF invoice, so selecting that alternative would not meet the stated acceptance check.

D:

Azure Vision Image Analysis supplies visual tags, captions, and supported detection results from image inputs; it may be used to describe visible objects and image content in uploaded photographs in a claims intake with mixed media, but the current requirement is instead to turn callers' recorded speech into text for a contact-center app. That separate capability does not prove that the deliverable is timestamped or plain speech transcription; It does not transcribe a telephone call or extract invoice line items into a domain-specific schema, so selecting that alternative would not meet the stated acceptance check.

E:

Document Intelligence analyzes document files with prebuilt or custom models for form-oriented field extraction; it may be used to extract recognized fields and values from invoices or receipts in a claims intake with mixed media, but the current requirement is instead to turn callers' recorded speech into text for a contact-center app. That separate capability does not prove that the deliverable is timestamped or plain speech transcription; It does not create natural-language conversation responses by itself, so selecting that alternative would not meet the stated acceptance check.

Question 29

Transcripts of medical-device identifiers are frequently wrong although common English is recognized correctly. What customization should be investigated?

  1. Poll or wait for the asynchronous video indexing job to complete before requesting final insights
  2. Add an explicit numeric claim-amount field to the analyzer extraction schema
  3. Train or adapt a supported Custom Speech model using relevant domain audio and text
  4. Route the multipage document to Document Intelligence rather than still-image captioning

Explanation

Correct Answer

C:

Domain-adapted speech recognition can improve specialized vocabulary when supported training data is representative. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after training or adapt a supported Custom Speech model using relevant domain audio and text.

Incorrect Answers

A:

Video indexing is asynchronous for many workflows; a pending job is not an empty transcript and should not be treated as finished in its own workflow, but that is a different correction from the present requirement to train or adapt a supported Custom Speech model using relevant domain audio and text. The reported issue would remain.

B:

Although useful elsewhere in the claims intake with mixed media, structured field definitions constrain the expected output contract more directly than asking for an unconstrained narrative; the present case is different and calls for the team to train or adapt a supported Custom Speech model using relevant domain audio and text.

D:

That change addresses another failure mode: document Intelligence has document and table extraction workflows; a single-image caption request is not a multipage parser; it does not address this case, for which the relevant intervention is to train or adapt a supported Custom Speech model using relevant domain audio and text.

Question 30

Reviewers must search interviews by spoken phrase, not merely by image tags. Which output should the application index?

  1. Inspect returned evidence or source grounding for that extracted date when available
  2. Enable a supported Read or OCR feature and consume its text spans
  3. Record and compare the extraction model or API version used for each processed document
  4. Index the transcript and its speech-derived timed segments from the video

Explanation

Correct Answer

D:

Speech transcripts support spoken-text search, while image tags describe visible scenes and cannot substitute for dialogue. In the claims intake with mixed media, confirm the fix by retesting the affected behavior after indexing the transcript and its speech-derived timed segments from the video.

Incorrect Answers

A:

That change addresses another failure mode: source evidence permits comparison of an inferred field with the underlying document or media segment instead of trusting an isolated JSON value; it does not address this case, for which the relevant intervention is to index the transcript and its speech-derived timed segments from the video.

B:

OCR returns recognized text content and location information where supported, while generic captions are not exact text transcripts in its own workflow, but that is a different correction from the present requirement to index the transcript and its speech-derived timed segments from the video. The reported issue would remain.

C:

Although useful elsewhere in the claims intake with mixed media, document extraction output can change across model releases; versioned provenance supports meaningful regression investigations; the present case is different and calls for the team to index the transcript and its speech-derived timed segments from the video.

Designing and Implementing a Microsoft Azure AI Solution (AI-102) Practice Tests