Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 38

Question 370

What test should engineers run after they apply approved provenance or watermark policy, rather than relying solely on a successful API response?

  1. Verify marking presence and metadata preservation through export steps
  2. Test malicious OCR payloads and confirm no unauthorized tool calls occur
  3. Audit unauthorized-mark incidence and reviewer decisions on test assets
  4. Measure retrieval success at the segment level with timestamp ground truth

Explanation

Correct Answer

A:

The accepted change is to apply approved provenance or watermark policy; this evaluation checks whether it can meet a disclosure requirement for generated advertising images where supported controls are available, since watermarking or provenance metadata can help identify synthetic assets, but does not replace editorial review or content safety checks.

Incorrect Answers

B:

This test concerns the ability to stop a photographed document from instructing an agent to disclose private data, for which instructions found by OCR inside external media are lower-trust content and must not gain tool authority; it does not establish whether the change to apply approved provenance or watermark policy works for the current release.

C:

This test concerns the ability to prevent generated marketing images from using prohibited marks or misrepresenting a sponsor, for which brand-policy review examines generated symbols and contextual use; generic harm classification does not establish trademark permission; it does not establish whether the change to apply approved provenance or watermark policy works for the current release.

D:

This test concerns the ability to classify chapters of a multi-topic video into retrievable units, for which video analyzers can emit scene segments, timestamps and structured or Markdown results that improve retrieval granularity; it does not establish whether the change to apply approved provenance or watermark policy works for the current release.

Question 371

The engineering change request is to use Azure Translator for text localization. What is the most direct acceptance criterion?

  1. Use human adequacy and terminology checks for representative translated instructions
  2. Evaluate correct action owners, due dates, and decision coverage on labeled calls
  3. Evaluate language identification accuracy on short and code-switched messages
  4. Compare meaning preservation and style adherence using bilingual human review

Explanation

Correct Answer

A:

The accepted change is to use Azure Translator for text localization; this evaluation checks whether it can translate stable product instructions into multiple supported target languages, since translator provides text translation with language configuration; domain terminology still requires quality review.

Incorrect Answers

B:

This test concerns the ability to create a handoff note that preserves commitments and unresolved action items from a long support call, for which conversation summaries must capture speakers, decisions and next steps rather than merely extract generic keywords; it does not establish whether the change to use Azure Translator for text localization works for the current release.

C:

This test concerns the ability to handle incoming tickets whose language varies and is not declared by the sender, for which language detection supplies a routing signal to a suitable translation or analysis path; short or mixed-language texts need caution; it does not establish whether the change to use Azure Translator for text localization works for the current release.

D:

This test concerns the ability to translate a conversational response while preserving a specified tone and contextual meaning, for which a multilingual model can use conversation context and style constraints; factual meaning and fluent tone need separate verification; it does not establish whether the change to use Azure Translator for text localization works for the current release.

Question 372

What test should engineers run after they extract key phrases for topic browsing, rather than relying solely on a successful API response?

  1. Score target-sentiment pair accuracy on annotated multi-aspect feedback
  2. Measure agreement with human sentiment labels, including mixed-polarity reviews
  3. Compare extracted topics with editor-labeled key concepts across articles
  4. Measure valid-schema rate and field-level extraction accuracy on labeled requests

Explanation

Correct Answer

C:

The accepted change is to extract key phrases for topic browsing; this evaluation checks whether it can surface the main subjects of short articles without requiring a fixed taxonomy, since key phrase extraction returns salient terms or phrases; it differs from assigning one of several predefined category labels.

Incorrect Answers

A:

This test concerns the ability to identify that a reviewer likes delivery speed but dislikes package quality in one comment, for which opinion mining links sentiment to specific targets or aspects instead of reporting only a document-level score; it does not establish whether the change to extract key phrases for topic browsing works for the current release.

B:

This test concerns the ability to detect broadly positive, negative or neutral attitudes expressed in reviews, for which sentiment analysis estimates evaluative polarity in text; it should not be used as a precise diagnosis of the reviewer's emotions; it does not establish whether the change to extract key phrases for topic browsing works for the current release.

D:

This test concerns the ability to extract account ID, reason category, and decision text into fixed machine-consumable fields, for which structured output controls field names and types for downstream code; domain validation must still check allowed values; it does not establish whether the change to extract key phrases for topic browsing works for the current release.

Question 373

After deployment, the team observes: The same regulatory phrase appears in three different translated forms across a document set. Which remediation addresses the likely cause?

  1. Add an approved term mapping through supported translation or post-editing controls
  2. Use a strict output schema and application validation for all expected fields
  3. Detect probable language and route to a supported pipeline with uncertainty handling
  4. Call the supported text translation service for each target locale and evaluate results

Explanation

Correct Answer

A:

The fault signal is that the same regulatory phrase appears in three different translated forms across a document set; the targeted fix works because glossary or controlled terminology guidance reduces inconsistent domain-language choices; blind literal translation may be unsafe, restoring the ability to preserve an approved glossary of medical or legal terms across translated responses.

Incorrect Answers

B:

If the failing function were to extract account ID, reason category, and decision text into fixed machine-consumable fields, this change could be relevant because structured output controls field names and types for downstream code; domain validation must still check allowed values; the actual symptom is that the same regulatory phrase appears in three different translated forms across a document set, which calls for a different repair.

C:

If the failing function were to handle incoming tickets whose language varies and is not declared by the sender, this change could be relevant because language detection supplies a routing signal to a suitable translation or analysis path; short or mixed-language texts need caution; the actual symptom is that the same regulatory phrase appears in three different translated forms across a document set, which calls for a different repair.

D:

If the failing function were to translate stable product instructions into multiple supported target languages, this change could be relevant because translator provides text translation with language configuration; domain terminology still requires quality review; the actual symptom is that the same regulatory phrase appears in three different translated forms across a document set, which calls for a different repair.

Question 374

In an architecture review, the key requirement is to translate stable product instructions into multiple supported target languages. Which solution directly satisfies it?

  1. Apply terminology constraints to regulated translations
  2. Use Azure Translator for text localization
  3. Detect source language before routing text
  4. Extract key phrases for topic browsing
  5. Redact private details from summarized text

Explanation

Correct Answer

B:

For the stated requirement to translate stable product instructions into multiple supported target languages, translator provides text translation with language configuration; domain terminology still requires quality review; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to preserve an approved glossary of medical or legal terms across translated responses, glossary or controlled terminology guidance reduces inconsistent domain-language choices; blind literal translation may be unsafe; this does not directly meet the required outcome to translate stable product instructions into multiple supported target languages.

C:

For a system whose objective is to handle incoming tickets whose language varies and is not declared by the sender, language detection supplies a routing signal to a suitable translation or analysis path; short or mixed-language texts need caution; this does not directly meet the required outcome to translate stable product instructions into multiple supported target languages.

D:

For a system whose objective is to surface the main subjects of short articles without requiring a fixed taxonomy, key phrase extraction returns salient terms or phrases; it differs from assigning one of several predefined category labels; this does not directly meet the required outcome to translate stable product instructions into multiple supported target languages.

E:

For a system whose objective is to produce a useful incident synopsis without retaining unnecessary personal identifiers, a redaction stage can remove sensitive spans from the generated artifact; summarization alone does not guarantee privacy; this does not directly meet the required outcome to translate stable product instructions into multiple supported target languages.

Question 375

An Azure AI workload passes basic health probes but has this issue: A support summary includes full customer phone numbers despite a policy against storing them. Which fix targets it?

  1. Detect probable language and route to a supported pipeline with uncertainty handling
  2. Supply precise domain definitions and representative labeled edge cases
  3. Use a strict output schema and application validation for all expected fields
  4. Apply PII detection and redact or tokenize the detected fields before logging

Explanation

Correct Answer

D:

The fault signal is that a support summary includes full customer phone numbers despite a policy against storing them; the targeted fix works because PII recognition identifies supported sensitive spans for redaction or controlled processing; keeping the raw text is a separate policy decision, restoring the ability to find private identifiers in user-provided text so they can be minimized or masked.

Incorrect Answers

A:

If the failing function were to handle incoming tickets whose language varies and is not declared by the sender, this change could be relevant because language detection supplies a routing signal to a suitable translation or analysis path; short or mixed-language texts need caution; the actual symptom is that a support summary includes full customer phone numbers despite a policy against storing them, which calls for a different repair.

B:

If the failing function were to extract a specialized compliance category when generic terminology is overloaded, this change could be relevant because task-specific definitions and examples help a model map ambiguous words to the intended domain meaning; the actual symptom is that a support summary includes full customer phone numbers despite a policy against storing them, which calls for a different repair.

C:

If the failing function were to extract account ID, reason category, and decision text into fixed machine-consumable fields, this change could be relevant because structured output controls field names and types for downstream code; domain validation must still check allowed values; the actual symptom is that a support summary includes full customer phone numbers despite a policy against storing them, which calls for a different repair.

Question 376

The acceptance criteria for a Foundry-based application require it to extract a specialized compliance category when generic terminology is overloaded. Which design should be approved?

  1. Classify messages against a defined taxonomy
  2. Request schema-constrained entity JSON
  3. Analyze sentiment in customer feedback
  4. Use domain examples to disambiguate terms

Explanation

Correct Answer

D:

For the stated requirement to extract a specialized compliance category when generic terminology is overloaded, task-specific definitions and examples help a model map ambiguous words to the intended domain meaning; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to route written requests into one of several approved service categories, a classifier assigns labels from an agreed taxonomy; named entity extraction alone does not decide message intent; this does not directly meet the required outcome to extract a specialized compliance category when generic terminology is overloaded.

B:

For a system whose objective is to extract account ID, reason category, and decision text into fixed machine-consumable fields, structured output controls field names and types for downstream code; domain validation must still check allowed values; this does not directly meet the required outcome to extract a specialized compliance category when generic terminology is overloaded.

C:

For a system whose objective is to detect broadly positive, negative or neutral attitudes expressed in reviews, sentiment analysis estimates evaluative polarity in text; it should not be used as a precise diagnosis of the reviewer's emotions; this does not directly meet the required outcome to extract a specialized compliance category when generic terminology is overloaded.

Question 377

After deployment, the team observes: Incident summaries still contain raw email addresses copied from source messages. Which remediation addresses the likely cause?

  1. Call the supported text translation service for each target locale and evaluate results
  2. Apply supported text content-safety checks and enforce category-specific handling
  3. Identify sensitive spans and sanitize source and output under the retention policy
  4. Use a source-constrained summary prompt and validate coverage of required clauses

Explanation

Correct Answer

C:

The fault signal is that incident summaries still contain raw email addresses copied from source messages; the targeted fix works because a redaction stage can remove sensitive spans from the generated artifact; summarization alone does not guarantee privacy, restoring the ability to produce a useful incident synopsis without retaining unnecessary personal identifiers.

Incorrect Answers

A:

If the failing function were to translate stable product instructions into multiple supported target languages, this change could be relevant because translator provides text translation with language configuration; domain terminology still requires quality review; the actual symptom is that incident summaries still contain raw email addresses copied from source messages, which calls for a different repair.

B:

If the failing function were to review user messages for policy-relevant hate, violence or self-harm content before generation, this change could be relevant because text safety classification identifies supported content categories; moderation decisions depend on configured policy and context; the actual symptom is that incident summaries still contain raw email addresses copied from source messages, which calls for a different repair.

D:

If the failing function were to condense a long policy document while preserving key exceptions and source support, this change could be relevant because grounded summarization uses supplied text as its evidence; a fluent summary must not introduce unsupported policy claims; the actual symptom is that incident summaries still contain raw email addresses copied from source messages, which calls for a different repair.

Question 378

An Azure AI workload passes basic health probes but has this issue: An automated call note lists product names but forgets the customer's promised refund date. Which fix targets it?

  1. Identify sensitive spans and sanitize source and output under the retention policy
  2. Use a source-constrained summary prompt and validate coverage of required clauses
  3. Translate with context and explicit tone requirements, then assess semantic fidelity
  4. Apply structured conversation summarization with action-item and attribution requirements

Explanation

Correct Answer

D:

The fault signal is that an automated call note lists product names but forgets the customer's promised refund date; the targeted fix works because conversation summaries must capture speakers, decisions and next steps rather than merely extract generic keywords, restoring the ability to create a handoff note that preserves commitments and unresolved action items from a long support call.

Incorrect Answers

A:

If the failing function were to produce a useful incident synopsis without retaining unnecessary personal identifiers, this change could be relevant because a redaction stage can remove sensitive spans from the generated artifact; summarization alone does not guarantee privacy; the actual symptom is that an automated call note lists product names but forgets the customer's promised refund date, which calls for a different repair.

B:

If the failing function were to condense a long policy document while preserving key exceptions and source support, this change could be relevant because grounded summarization uses supplied text as its evidence; a fluent summary must not introduce unsupported policy claims; the actual symptom is that an automated call note lists product names but forgets the customer's promised refund date, which calls for a different repair.

C:

If the failing function were to translate a conversational response while preserving a specified tone and contextual meaning, this change could be relevant because a multilingual model can use conversation context and style constraints; factual meaning and fluent tone need separate verification; the actual symptom is that an automated call note lists product names but forgets the customer's promised refund date, which calls for a different repair.

Question 379

After implementing a workflow to detect source language before routing text, the team needs a robust release gate. Which check should it choose?

  1. Measure terminology consistency and reviewer corrections across languages
  2. Use human adequacy and terminology checks for representative translated instructions
  3. Measure entity-level precision and recall with annotated names and locations
  4. Evaluate language identification accuracy on short and code-switched messages

Explanation

Correct Answer

D:

The accepted change is to detect source language before routing text; this evaluation checks whether it can handle incoming tickets whose language varies and is not declared by the sender, since language detection supplies a routing signal to a suitable translation or analysis path; short or mixed-language texts need caution.

Incorrect Answers

A:

This test concerns the ability to preserve an approved glossary of medical or legal terms across translated responses, for which glossary or controlled terminology guidance reduces inconsistent domain-language choices; blind literal translation may be unsafe; it does not establish whether the change to detect source language before routing text works for the current release.

B:

This test concerns the ability to translate stable product instructions into multiple supported target languages, for which translator provides text translation with language configuration; domain terminology still requires quality review; it does not establish whether the change to detect source language before routing text works for the current release.

C:

This test concerns the ability to identify the people, organizations and places mentioned in incident narratives, for which named entity recognition identifies and classifies mentions in text; it does not by itself establish who authored the document or whether claims are true; it does not establish whether the change to detect source language before routing text works for the current release.

Question 380

An operator can reproduce this error: An article portal shows generic word frequency rather than meaningful topics. Which targeted repair is justified?

  1. Apply entity extraction and review entity spans and categories in sample reports
  2. Identify sensitive spans and sanitize source and output under the retention policy
  3. Apply sentiment analysis with representative domain testing and interpret mixed cases carefully
  4. Use key phrase extraction and filter noisy or redundant phrases

Explanation

Correct Answer

D:

The fault signal is that an article portal shows generic word frequency rather than meaningful topics; the targeted fix works because key phrase extraction returns salient terms or phrases; it differs from assigning one of several predefined category labels, restoring the ability to surface the main subjects of short articles without requiring a fixed taxonomy.

Incorrect Answers

A:

If the failing function were to identify the people, organizations and places mentioned in incident narratives, this change could be relevant because named entity recognition identifies and classifies mentions in text; it does not by itself establish who authored the document or whether claims are true; the actual symptom is that an article portal shows generic word frequency rather than meaningful topics, which calls for a different repair.

B:

If the failing function were to produce a useful incident synopsis without retaining unnecessary personal identifiers, this change could be relevant because a redaction stage can remove sensitive spans from the generated artifact; summarization alone does not guarantee privacy; the actual symptom is that an article portal shows generic word frequency rather than meaningful topics, which calls for a different repair.

C:

If the failing function were to detect broadly positive, negative or neutral attitudes expressed in reviews, this change could be relevant because sentiment analysis estimates evaluative polarity in text; it should not be used as a precise diagnosis of the reviewer's emotions; the actual symptom is that an article portal shows generic word frequency rather than meaningful topics, which calls for a different repair.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests