Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 29

Question 280

The support ticket describes this behavior: A high-stakes assistant guesses an exception code when no authoritative document contains it. Which engineering change should come first?

  1. Assess false positives with a labeled set and tune supported filter settings under policy approval
  2. Classify task complexity and route easy cases to a validated lower-cost model
  3. Summarize completed threads and retain exact active identifiers and constraints
  4. Implement a supported no-evidence response and escalation route for low-support cases

Explanation

Correct Answer

D:

The fault signal is that a high-stakes assistant guesses an exception code when no authoritative document contains it; the targeted fix works because a calibrated no-answer path is safer than fabricating a missing fact; thresholds should be tuned on labeled questions, restoring the ability to prevent confident unsupported answers when a retrieval score or evidence threshold is weak.

Incorrect Answers

A:

If the failing function were to reduce legitimate request blocks without weakening essential guardrails, this change could be relevant because precision and recall of safety classifications require policy-labeled evaluation rather than a blanket filter disablement; the actual symptom is that a high-stakes assistant guesses an exception code when no authoritative document contains it, which calls for a different repair.

B:

If the failing function were to send bounded simple queries to economical models and harder reasoning tasks to stronger ones, this change could be relevant because a routing policy can select among models based on evaluated task difficulty and quality constraints; the actual symptom is that a high-stakes assistant guesses an exception code when no authoritative document contains it, which calls for a different repair.

C:

If the failing function were to reduce token cost in long sessions while keeping current commitments and unresolved tasks, this change could be relevant because selective compression retains required state but excludes repeated old messages; simply raising token limits does not fix unnecessary spend; the actual symptom is that a high-stakes assistant guesses an exception code when no authoritative document contains it, which calls for a different repair.

Question 281

The implementation does not meet its business requirement: Every trivial FAQ query uses a costly reasoning model even though a small model performs equivalently. Which corrective action is appropriate?

  1. Tie cached contexts to source revision and invalidate on indexed-content changes
  2. Run a business-rule validator after schema checks and reject impossible state transitions
  3. Classify task complexity and route easy cases to a validated lower-cost model
  4. Summarize completed threads and retain exact active identifiers and constraints

Explanation

Correct Answer

C:

The fault signal is that every trivial FAQ query uses a costly reasoning model even though a small model performs equivalently; the targeted fix works because a routing policy can select among models based on evaluated task difficulty and quality constraints, restoring the ability to send bounded simple queries to economical models and harder reasoning tasks to stronger ones.

Incorrect Answers

A:

If the failing function were to avoid repeated retrieval for unchanged documents while preventing stale grounding, this change could be relevant because versioned cache keys or invalidation preserve freshness when approved source content changes; the actual symptom is that every trivial FAQ query uses a costly reasoning model even though a small model performs equivalently, which calls for a different repair.

B:

If the failing function were to catch schema-valid yet impossible business outputs before a workflow proceeds, this change could be relevant because a type-correct JSON value may still violate business invariants; deterministic post-validation enforces the rules; the actual symptom is that every trivial FAQ query uses a costly reasoning model even though a small model performs equivalently, which calls for a different repair.

D:

If the failing function were to reduce token cost in long sessions while keeping current commitments and unresolved tasks, this change could be relevant because selective compression retains required state but excludes repeated old messages; simply raising token limits does not fix unnecessary spend; the actual symptom is that every trivial FAQ query uses a costly reasoning model even though a small model performs equivalently, which calls for a different repair.

Question 282

A solution architect is deciding how to lower overhead when indexing many small documents without changing semantic meaning. Which Azure AI pattern addresses that task?

  1. Batch independent embeddings for ingestion efficiency
  2. Review content-filter trade-offs on representative examples
  3. Cache retrieved context with source versioning
  4. Use uncertainty-aware abstention

Explanation

Correct Answer

A:

For the stated requirement to lower overhead when indexing many small documents without changing semantic meaning, batching compatible embedding requests can improve throughput within service rate limits; errors require granular retries; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to reduce legitimate request blocks without weakening essential guardrails, precision and recall of safety classifications require policy-labeled evaluation rather than a blanket filter disablement; this does not directly meet the required outcome to lower overhead when indexing many small documents without changing semantic meaning.

C:

For a system whose objective is to avoid repeated retrieval for unchanged documents while preventing stale grounding, versioned cache keys or invalidation preserve freshness when approved source content changes; this does not directly meet the required outcome to lower overhead when indexing many small documents without changing semantic meaning.

D:

For a system whose objective is to prevent confident unsupported answers when a retrieval score or evidence threshold is weak, a calibrated no-answer path is safer than fabricating a missing fact; thresholds should be tuned on labeled questions; this does not directly meet the required outcome to lower overhead when indexing many small documents without changing semantic meaning.

Question 283

An Azure AI delivery has 2 independently testable requirements: (1) detect whether changing instructions improves one use case while damaging another; and (2) avoid treating a model's agreement with its own answer as proof of correctness. Select the most directly specialized implementation for each requirement. Choose TWO.

  1. Cache retrieved context with source versioning
  2. Compare model candidates with task-specific benchmarks
  3. Add a critique-and-revise stage for draft quality
  4. Inspect prompt regressions using paired examples
  5. Use external evidence to verify generated claims

Explanation

Correct Answers

D:

For the stated requirement to detect whether changing instructions improves one use case while damaging another, paired cases give controlled evidence about prompt effects and expose class-specific regressions; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to detect whether changing instructions improves one use case while damaging another, while the other deliverables require separate capabilities.

E:

For the stated requirement to avoid treating a model's agreement with its own answer as proof of correctness, independent source retrieval or deterministic tools can verify claims beyond the generator's self-confidence; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to avoid treating a model's agreement with its own answer as proof of correctness, while the other deliverables require separate capabilities.

Incorrect Answers

A:

For a system whose objective is to avoid repeated retrieval for unchanged documents while preventing stale grounding, versioned cache keys or invalidation preserve freshness when approved source content changes; this does not directly meet the required outcome to detect whether changing instructions improves one use case while damaging another.

B:

For a system whose objective is to choose between smaller and larger models based on business quality and cost, representative evaluations reveal the trade-off among quality, latency and spend; model size is an imperfect proxy; this does not directly meet the required outcome to detect whether changing instructions improves one use case while damaging another.

C:

For a system whose objective is to improve an initial generated report by checking it against a defined factual and style rubric, a critique step can identify defects in an earlier draft when grounded in explicit criteria; self-review alone is not an independent fact source; this does not directly meet the required outcome to detect whether changing instructions improves one use case while damaging another.

Question 284

A production AI solution is being scoped for production. The non-negotiable task is to reduce token cost in long sessions while keeping current commitments and unresolved tasks. Which implementation fits?

  1. Lower sampling temperature for stable classifications
  2. Use model routing by task complexity
  3. Review content-filter trade-offs on representative examples
  4. Compare model candidates with task-specific benchmarks
  5. Use uncertainty-aware abstention
  6. Trim redundant conversation context

Explanation

Correct Answer

F:

For the stated requirement to reduce token cost in long sessions while keeping current commitments and unresolved tasks, selective compression retains required state but excludes repeated old messages; simply raising token limits does not fix unnecessary spend; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to reduce unwarranted variability when an output must follow a fixed decision rubric, lower temperature can reduce sampling randomness where supported; it does not guarantee factual accuracy or replace schema validation; this does not directly meet the required outcome to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

B:

For a system whose objective is to send bounded simple queries to economical models and harder reasoning tasks to stronger ones, a routing policy can select among models based on evaluated task difficulty and quality constraints; this does not directly meet the required outcome to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

C:

For a system whose objective is to reduce legitimate request blocks without weakening essential guardrails, precision and recall of safety classifications require policy-labeled evaluation rather than a blanket filter disablement; this does not directly meet the required outcome to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

D:

For a system whose objective is to choose between smaller and larger models based on business quality and cost, representative evaluations reveal the trade-off among quality, latency and spend; model size is an imperfect proxy; this does not directly meet the required outcome to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

E:

For a system whose objective is to prevent confident unsupported answers when a retrieval score or evidence threshold is weak, a calibrated no-answer path is safer than fabricating a missing fact; thresholds should be tuned on labeled questions; this does not directly meet the required outcome to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

Question 285

A product team must reduce legitimate request blocks without weakening essential guardrails. What is the most appropriate Azure AI implementation?

  1. Trim redundant conversation context
  2. Review content-filter trade-offs on representative examples
  3. Add a critique-and-revise stage for draft quality
  4. Use uncertainty-aware abstention

Explanation

Correct Answer

B:

For the stated requirement to reduce legitimate request blocks without weakening essential guardrails, precision and recall of safety classifications require policy-labeled evaluation rather than a blanket filter disablement; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to reduce token cost in long sessions while keeping current commitments and unresolved tasks, selective compression retains required state but excludes repeated old messages; simply raising token limits does not fix unnecessary spend; this does not directly meet the required outcome to reduce legitimate request blocks without weakening essential guardrails.

C:

For a system whose objective is to improve an initial generated report by checking it against a defined factual and style rubric, a critique step can identify defects in an earlier draft when grounded in explicit criteria; self-review alone is not an independent fact source; this does not directly meet the required outcome to reduce legitimate request blocks without weakening essential guardrails.

D:

For a system whose objective is to prevent confident unsupported answers when a retrieval score or evidence threshold is weak, a calibrated no-answer path is safer than fabricating a missing fact; thresholds should be tuned on labeled questions; this does not directly meet the required outcome to reduce legitimate request blocks without weakening essential guardrails.

Question 286

Production telemetry reveals this failure: A draft report misses required sections and contradicts a supplied figures table. Which change most directly corrects it?

  1. Cross-check each material claim against approved authoritative sources or executable rules
  2. Group independent embedding requests into supported batches with rate-aware retries
  3. Move enduring role constraints into trusted instruction channels and enforce permissions in code
  4. Compare fixed regression sets covering both summarization and citation tasks across prompt versions
  5. Evaluate the draft against a rubric and regenerate only the failing sections using source evidence

Explanation

Correct Answer

E:

The fault signal is that a draft report misses required sections and contradicts a supplied figures table; the targeted fix works because a critique step can identify defects in an earlier draft when grounded in explicit criteria; self-review alone is not an independent fact source, restoring the ability to improve an initial generated report by checking it against a defined factual and style rubric.

Incorrect Answers

A:

If the failing function were to avoid treating a model's agreement with its own answer as proof of correctness, this change could be relevant because independent source retrieval or deterministic tools can verify claims beyond the generator's self-confidence; the actual symptom is that a draft report misses required sections and contradicts a supplied figures table, which calls for a different repair.

B:

If the failing function were to lower overhead when indexing many small documents without changing semantic meaning, this change could be relevant because batching compatible embedding requests can improve throughput within service rate limits; errors require granular retries; the actual symptom is that a draft report misses required sections and contradicts a supplied figures table, which calls for a different repair.

C:

If the failing function were to keep permanent application behavior separate from ordinary user requests, this change could be relevant because high-priority instructions establish the app's role and safety boundaries, but external authorization still requires programmatic enforcement; the actual symptom is that a draft report misses required sections and contradicts a supplied figures table, which calls for a different repair.

D:

If the failing function were to detect whether changing instructions improves one use case while damaging another, this change could be relevant because paired cases give controlled evidence about prompt effects and expose class-specific regressions; the actual symptom is that a draft report misses required sections and contradicts a supplied figures table, which calls for a different repair.

Question 287

After deployment, the team observes: The model receives a growing transcript of resolved issues and input-token cost doubles each hour. Which remediation addresses the likely cause?

  1. Run the same labeled task set across candidate models and compare operational metrics
  2. Summarize completed threads and retain exact active identifiers and constraints
  3. Group independent embedding requests into supported batches with rate-aware retries
  4. Assess false positives with a labeled set and tune supported filter settings under policy approval

Explanation

Correct Answer

B:

The fault signal is that the model receives a growing transcript of resolved issues and input-token cost doubles each hour; the targeted fix works because selective compression retains required state but excludes repeated old messages; simply raising token limits does not fix unnecessary spend, restoring the ability to reduce token cost in long sessions while keeping current commitments and unresolved tasks.

Incorrect Answers

A:

If the failing function were to choose between smaller and larger models based on business quality and cost, this change could be relevant because representative evaluations reveal the trade-off among quality, latency and spend; model size is an imperfect proxy; the actual symptom is that the model receives a growing transcript of resolved issues and input-token cost doubles each hour, which calls for a different repair.

C:

If the failing function were to lower overhead when indexing many small documents without changing semantic meaning, this change could be relevant because batching compatible embedding requests can improve throughput within service rate limits; errors require granular retries; the actual symptom is that the model receives a growing transcript of resolved issues and input-token cost doubles each hour, which calls for a different repair.

D:

If the failing function were to reduce legitimate request blocks without weakening essential guardrails, this change could be relevant because precision and recall of safety classifications require policy-labeled evaluation rather than a blanket filter disablement; the actual symptom is that the model receives a growing transcript of resolved issues and input-token cost doubles each hour, which calls for a different repair.

Question 288

The implementation does not meet its business requirement: The model confuses two internal categories that differ by a subtle phrasing rule. Which corrective action is appropriate?

  1. Move enduring role constraints into trusted instruction channels and enforce permissions in code
  2. Implement a supported no-evidence response and escalation route for low-support cases
  3. Add a small diverse set of labeled examples with boundary cases to the prompt
  4. Reduce stochastic sampling and validate the mapping on labeled repeated tests

Explanation

Correct Answer

C:

The fault signal is that the model confuses two internal categories that differ by a subtle phrasing rule; the targeted fix works because well-chosen examples illustrate the desired input-output mapping, including ambiguous edge cases; they must be representative, restoring the ability to teach a model a specialized labeling convention that general instructions describe poorly.

Incorrect Answers

A:

If the failing function were to keep permanent application behavior separate from ordinary user requests, this change could be relevant because high-priority instructions establish the app's role and safety boundaries, but external authorization still requires programmatic enforcement; the actual symptom is that the model confuses two internal categories that differ by a subtle phrasing rule, which calls for a different repair.

B:

If the failing function were to prevent confident unsupported answers when a retrieval score or evidence threshold is weak, this change could be relevant because a calibrated no-answer path is safer than fabricating a missing fact; thresholds should be tuned on labeled questions; the actual symptom is that the model confuses two internal categories that differ by a subtle phrasing rule, which calls for a different repair.

D:

If the failing function were to reduce unwarranted variability when an output must follow a fixed decision rubric, this change could be relevant because lower temperature can reduce sampling randomness where supported; it does not guarantee factual accuracy or replace schema validation; the actual symptom is that the model confuses two internal categories that differ by a subtle phrasing rule, which calls for a different repair.

Question 289

The team plans to add a critique-and-revise stage for draft quality. Which test provides the clearest evidence that this change works?

  1. Measure indexed chunks per minute, per-item failures, and sustained quota usage
  2. Report per-category deltas and inspect failed examples instead of relying on one aggregate score
  3. Measure claims corroborated by independent evidence and unresolved discrepancies
  4. Measure rubric violations and unsupported statements before and after revision

Explanation

Correct Answer

D:

The accepted change is to add a critique-and-revise stage for draft quality; this evaluation checks whether it can improve an initial generated report by checking it against a defined factual and style rubric, since a critique step can identify defects in an earlier draft when grounded in explicit criteria; self-review alone is not an independent fact source.

Incorrect Answers

A:

This test concerns the ability to lower overhead when indexing many small documents without changing semantic meaning, for which batching compatible embedding requests can improve throughput within service rate limits; errors require granular retries; it does not establish whether the change to add a critique-and-revise stage for draft quality works for the current release.

B:

This test concerns the ability to detect whether changing instructions improves one use case while damaging another, for which paired cases give controlled evidence about prompt effects and expose class-specific regressions; it does not establish whether the change to add a critique-and-revise stage for draft quality works for the current release.

C:

This test concerns the ability to avoid treating a model's agreement with its own answer as proof of correctness, for which independent source retrieval or deterministic tools can verify claims beyond the generator's self-confidence; it does not establish whether the change to add a critique-and-revise stage for draft quality works for the current release.

Question 290

An architecture board is assessing 2 separate deliverables: (1) avoid long generated replies when a client requires a short summary or fixed output budget; and (2) keep permanent application behavior separate from ordinary user requests. Which distinct implementations meet these deliverables? Choose TWO.

  1. Constrain response length for concise summaries
  2. Use explicit system and developer instructions for policy
  3. Lower sampling temperature for stable classifications
  4. Use explicit few-shot examples for format nuance
  5. Batch independent embeddings for ingestion efficiency
  6. Cache retrieved context with source versioning

Explanation

Correct Answers

A:

For the stated requirement to avoid long generated replies when a client requires a short summary or fixed output budget, a sensible output-token cap and concise instruction bound generation size, but a cap that is too small may truncate content; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to avoid long generated replies when a client requires a short summary or fixed output budget, while the other deliverables require separate capabilities.

B:

For the stated requirement to keep permanent application behavior separate from ordinary user requests, high-priority instructions establish the app's role and safety boundaries, but external authorization still requires programmatic enforcement; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to keep permanent application behavior separate from ordinary user requests, while the other deliverables require separate capabilities.

Incorrect Answers

C:

For a system whose objective is to reduce unwarranted variability when an output must follow a fixed decision rubric, lower temperature can reduce sampling randomness where supported; it does not guarantee factual accuracy or replace schema validation; this does not directly meet the required outcome to keep permanent application behavior separate from ordinary user requests.

D:

For a system whose objective is to teach a model a specialized labeling convention that general instructions describe poorly, well-chosen examples illustrate the desired input-output mapping, including ambiguous edge cases; they must be representative; this does not directly meet the required outcome to keep permanent application behavior separate from ordinary user requests.

E:

For a system whose objective is to lower overhead when indexing many small documents without changing semantic meaning, batching compatible embedding requests can improve throughput within service rate limits; errors require granular retries; this does not directly meet the required outcome to keep permanent application behavior separate from ordinary user requests.

F:

For a system whose objective is to avoid repeated retrieval for unchanged documents while preventing stale grounding, versioned cache keys or invalidation preserve freshness when approved source content changes; this does not directly meet the required outcome to keep permanent application behavior separate from ordinary user requests.

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