Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 19

Question 180

A product team needs its AI workflow to let a compliance reviewer locate the specific evidence supporting each generated assertion. What technical approach most directly enables that?

  1. Use passage-level citations in RAG answers
  2. Use constrained structured JSON output
  3. Implement fallback when retrieval is insufficient
  4. Use chunk overlap for boundary-spanning facts

Explanation

Correct Answer

A:

For the stated requirement to let a compliance reviewer locate the specific evidence supporting each generated assertion, citations tied to retrieved chunks improve traceability; a source list without claim alignment does not prove each claim; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to send model-produced classifications into a strict downstream API contract, schema-constrained generation helps enforce machine-readable fields and types; application-level validation is still needed for business rules; this does not directly meet the required outcome to let a compliance reviewer locate the specific evidence supporting each generated assertion.

C:

For a system whose objective is to avoid fabricating an answer to a question not covered by the organization's knowledge collection, an evidence threshold and abstention path distinguish unknown facts from model completion skill; confident text is not proof of source support; this does not directly meet the required outcome to let a compliance reviewer locate the specific evidence supporting each generated assertion.

D:

For a system whose objective is to keep context that crosses page or section boundaries from being lost during indexing, chunk overlap can retain nearby context across splits, although excessive overlap inflates index size and duplicates retrieval; this does not directly meet the required outcome to let a compliance reviewer locate the specific evidence supporting each generated assertion.

Question 181

An operator can reproduce this error: The assistant cites an old procedure even though updated instructions are available in the search corpus. Which targeted repair is justified?

  1. Detect empty or low-confidence retrieval and return a supported no-answer or escalation response
  2. Retrieve current authorized passages and cite them when forming the response
  3. Request structured outputs conforming to the expected schema and validate before ingestion
  4. Provide image content to a vision-capable model with a bounded question and verification rubric

Explanation

Correct Answer

B:

The fault signal is that the assistant cites an old procedure even though updated instructions are available in the search corpus; the targeted fix works because RAG retrieves relevant evidence at runtime and gives it to a generative model; model pretraining alone cannot guarantee current policy text, restoring the ability to answer questions about changing internal procedures with source-backed responses.

Incorrect Answers

A:

If the failing function were to avoid fabricating an answer to a question not covered by the organization's knowledge collection, this change could be relevant because an evidence threshold and abstention path distinguish unknown facts from model completion skill; confident text is not proof of source support; the actual symptom is that the assistant cites an old procedure even though updated instructions are available in the search corpus, which calls for a different repair.

C:

If the failing function were to send model-produced classifications into a strict downstream API contract, this change could be relevant because schema-constrained generation helps enforce machine-readable fields and types; application-level validation is still needed for business rules; the actual symptom is that the assistant cites an old procedure even though updated instructions are available in the search corpus, which calls for a different repair.

D:

If the failing function were to reason about a photo and related user instructions together in one supported inference flow, this change could be relevant because a multimodal model can interpret visual input alongside textual context; relying on OCR alone loses non-textual relationships; the actual symptom is that the assistant cites an old procedure even though updated instructions are available in the search corpus, which calls for a different repair.

Question 182

The team plans to run independent tool requests concurrently where safe. Which test provides the clearest evidence that this change works?

  1. Verify multi-hop task completion and loop termination under adversarial inputs
  2. Trace required step ordering and block completions that omit a mandatory stage
  3. Count handled failures and confirm consistent error contracts on negative test suites
  4. Compare critical-path latency and result consistency with serial execution

Explanation

Correct Answer

D:

The accepted change is to run independent tool requests concurrently where safe; this evaluation checks whether it can reduce latency when separate read-only lookups have no ordering dependency, since parallelizing independent operations shortens critical-path latency but is unsafe when one action requires another's output.

Incorrect Answers

A:

This test concerns the ability to allow a model to request multiple independent data lookups while limiting cost and duration, for which an orchestration loop can inspect tool requests, execute authorized tools, return results, and terminate under explicit limits; it does not establish whether the change to run independent tool requests concurrently where safe works for the current release.

B:

This test concerns the ability to execute a known approval workflow where inputs must be validated before a model drafts a response, for which explicit workflow stages make ordering and failure handling reproducible; free-form agent planning is unnecessary for a fixed process; it does not establish whether the change to run independent tool requests concurrently where safe works for the current release.

C:

This test concerns the ability to verify that an app recovers safely from empty messages, long context, and invalid tool responses, for which boundary-case testing surfaces parsing and orchestration failures that standard benchmark prompts miss; it does not establish whether the change to run independent tool requests concurrently where safe works for the current release.

Question 183

A product team must reason about a photo and related user instructions together in one supported inference flow. What is the most appropriate Azure AI implementation?

  1. Use a multimodal request for document-plus-image reasoning
  2. Use constrained structured JSON output
  3. Use grounded retrieval-augmented generation
  4. Use token-based credentials rather than embedded keys

Explanation

Correct Answer

A:

For the stated requirement to reason about a photo and related user instructions together in one supported inference flow, a multimodal model can interpret visual input alongside textual context; relying on OCR alone loses non-textual relationships; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to send model-produced classifications into a strict downstream API contract, schema-constrained generation helps enforce machine-readable fields and types; application-level validation is still needed for business rules; this does not directly meet the required outcome to reason about a photo and related user instructions together in one supported inference flow.

C:

For a system whose objective is to answer questions about changing internal procedures with source-backed responses, RAG retrieves relevant evidence at runtime and gives it to a generative model; model pretraining alone cannot guarantee current policy text; this does not directly meet the required outcome to reason about a photo and related user instructions together in one supported inference flow.

D:

For a system whose objective is to authenticate a server-side generative app through Entra where supported, a workload identity and appropriate RBAC reduce static credential exposure while preserving auditable authorization decisions; this does not directly meet the required outcome to reason about a photo and related user instructions together in one supported inference flow.

Question 184

An operator can reproduce this error: A benefits assistant blends expired and current policies from multiple departments. Which targeted repair is justified?

  1. Request structured outputs conforming to the expected schema and validate before ingestion
  2. Filter authorized indexed chunks using tenant, department, and effective-date fields
  3. Adjust chunk boundaries and limited overlap, then rebuild affected embeddings
  4. Attach retrieved chunk references to the claims they substantiate and validate that mapping

Explanation

Correct Answer

B:

The fault signal is that a benefits assistant blends expired and current policies from multiple departments; the targeted fix works because retrieval filters narrow the evidence set by structured attributes before the model sees it; prompt instructions are not security filters, restoring the ability to ensure answers include only documents for a user's department and effective policy year.

Incorrect Answers

A:

If the failing function were to send model-produced classifications into a strict downstream API contract, this change could be relevant because schema-constrained generation helps enforce machine-readable fields and types; application-level validation is still needed for business rules; the actual symptom is that a benefits assistant blends expired and current policies from multiple departments, which calls for a different repair.

C:

If the failing function were to keep context that crosses page or section boundaries from being lost during indexing, this change could be relevant because chunk overlap can retain nearby context across splits, although excessive overlap inflates index size and duplicates retrieval; the actual symptom is that a benefits assistant blends expired and current policies from multiple departments, which calls for a different repair.

D:

If the failing function were to let a compliance reviewer locate the specific evidence supporting each generated assertion, this change could be relevant because citations tied to retrieved chunks improve traceability; a source list without claim alignment does not prove each claim; the actual symptom is that a benefits assistant blends expired and current policies from multiple departments, which calls for a different repair.

Question 185

The implementation does not meet its business requirement: The assistant quotes a true paragraph about benefits but ignores the user's eligibility question. Which corrective action is appropriate?

  1. Run A/B offline comparisons on a representative fixed dataset before rollout
  2. Add question-answer relevance evaluations with labeled examples and failure review
  3. Evaluate claim support against the supplied passages and require repair or abstention on failures
  4. Provide image content to a vision-capable model with a bounded question and verification rubric

Explanation

Correct Answer

B:

The fault signal is that the assistant quotes a true paragraph about benefits but ignores the user's eligibility question; the targeted fix works because relevance evaluates how directly a response serves the request; factual grounding and response usefulness are related but distinct axes, restoring the ability to detect answers that are grounded in retrieved text but do not address the actual question.

Incorrect Answers

A:

If the failing function were to choose between candidate prompt versions by measured task quality rather than intuition, this change could be relevant because paired evaluation on identical cases isolates prompt impact; production traffic shifts make uncontrolled comparisons misleading; the actual symptom is that the assistant quotes a true paragraph about benefits but ignores the user's eligibility question, which calls for a different repair.

C:

If the failing function were to catch plausible statements that are not substantiated by the retrieved evidence, this change could be relevant because faithfulness checks whether generated claims follow provided source material, not whether wording is eloquent; the actual symptom is that the assistant quotes a true paragraph about benefits but ignores the user's eligibility question, which calls for a different repair.

D:

If the failing function were to reason about a photo and related user instructions together in one supported inference flow, this change could be relevant because a multimodal model can interpret visual input alongside textual context; relying on OCR alone loses non-textual relationships; the actual symptom is that the assistant quotes a true paragraph about benefits but ignores the user's eligibility question, which calls for a different repair.

Question 186

What test should engineers run after they add bounded tool-use loops for multi-step tasks, rather than relying solely on a successful API response?

  1. Inspect request-target metadata and run an environment-isolation smoke test
  2. Compare branch decisions against a complete table of allowed status transitions
  3. Trace required step ordering and block completions that omit a mandatory stage
  4. Verify multi-hop task completion and loop termination under adversarial inputs

Explanation

Correct Answer

D:

The accepted change is to add bounded tool-use loops for multi-step tasks; this evaluation checks whether it can allow a model to request multiple independent data lookups while limiting cost and duration, since an orchestration loop can inspect tool requests, execute authorized tools, return results, and terminate under explicit limits.

Incorrect Answers

A:

This test concerns the ability to avoid routing production questions to a test model through hardcoded SDK settings, for which endpoint and deployment identifiers are environment-specific; runtime configuration keeps inference calls aligned to the project; it does not establish whether the change to add bounded tool-use loops for multi-step tasks works for the current release.

B:

This test concerns the ability to choose among distinct business flows based on a validated account status, for which rules engines offer predictable decisions for simple structured conditions, while language models handle unstructured interpretation; it does not establish whether the change to add bounded tool-use loops for multi-step tasks works for the current release.

C:

This test concerns the ability to execute a known approval workflow where inputs must be validated before a model drafts a response, for which explicit workflow stages make ordering and failure handling reproducible; free-form agent planning is unnecessary for a fixed process; it does not establish whether the change to add bounded tool-use loops for multi-step tasks works for the current release.

Question 187

An engineer receives this incident report: The application builds with an API key committed to a public repository. What should be changed?

  1. Evaluate claim support against the supplied passages and require repair or abstention on failures
  2. Configure a credential chain using the workload identity and least-privilege role grants
  3. Load approved project and deployment settings from the production environment
  4. Implement a controlled request-tool-result loop with a maximum number of steps
  5. Reject invalid parameters with application-side validation before the business API call

Explanation

Correct Answer

B:

The fault signal is that the application builds with an API key committed to a public repository; the targeted fix works because a workload identity and appropriate RBAC reduce static credential exposure while preserving auditable authorization decisions, restoring the ability to authenticate a server-side generative app through Entra where supported.

Incorrect Answers

A:

If the failing function were to catch plausible statements that are not substantiated by the retrieved evidence, this change could be relevant because faithfulness checks whether generated claims follow provided source material, not whether wording is eloquent; the actual symptom is that the application builds with an API key committed to a public repository, which calls for a different repair.

C:

If the failing function were to avoid routing production questions to a test model through hardcoded SDK settings, this change could be relevant because endpoint and deployment identifiers are environment-specific; runtime configuration keeps inference calls aligned to the project; the actual symptom is that the application builds with an API key committed to a public repository, which calls for a different repair.

D:

If the failing function were to allow a model to request multiple independent data lookups while limiting cost and duration, this change could be relevant because an orchestration loop can inspect tool requests, execute authorized tools, return results, and terminate under explicit limits; the actual symptom is that the application builds with an API key committed to a public repository, which calls for a different repair.

E:

If the failing function were to prevent malformed model-generated parameters from reaching a write-capable API, this change could be relevant because tool schemas guide argument shape, but the application must check types, constraints, identity and authorization before execution; the actual symptom is that the application builds with an API key committed to a public repository, which calls for a different repair.

Question 188

The planned deployment includes a mechanism to test application robustness on malformed inputs. Which outcome should the QA team measure?

  1. Count handled failures and confirm consistent error contracts on negative test suites
  2. Test for cross-department leakage and expired-document citation rate
  3. Compare win rate, task accuracy, and safety failures with statistical uncertainty
  4. Compute unsupported-claim rate on source-grounded summaries

Explanation

Correct Answer

A:

The accepted change is to test application robustness on malformed inputs; this evaluation checks whether it can verify that an app recovers safely from empty messages, long context, and invalid tool responses, since boundary-case testing surfaces parsing and orchestration failures that standard benchmark prompts miss.

Incorrect Answers

B:

This test concerns the ability to ensure answers include only documents for a user's department and effective policy year, for which retrieval filters narrow the evidence set by structured attributes before the model sees it; prompt instructions are not security filters; it does not establish whether the change to test application robustness on malformed inputs works for the current release.

C:

This test concerns the ability to choose between candidate prompt versions by measured task quality rather than intuition, for which paired evaluation on identical cases isolates prompt impact; production traffic shifts make uncontrolled comparisons misleading; it does not establish whether the change to test application robustness on malformed inputs works for the current release.

D:

This test concerns the ability to catch plausible statements that are not substantiated by the retrieved evidence, for which faithfulness checks whether generated claims follow provided source material, not whether wording is eloquent; it does not establish whether the change to test application robustness on malformed inputs works for the current release.

Question 189

A customer-facing AI service must consistently ensure answers include only documents for a user's department and effective policy year. Which architecture decision is most defensible?

  1. Route conditional branches using rules
  2. Apply retrieval metadata filters before generation
  3. Use chunk overlap for boundary-spanning facts
  4. Implement fallback when retrieval is insufficient

Explanation

Correct Answer

B:

For the stated requirement to ensure answers include only documents for a user's department and effective policy year, retrieval filters narrow the evidence set by structured attributes before the model sees it; prompt instructions are not security filters; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to choose among distinct business flows based on a validated account status, rules engines offer predictable decisions for simple structured conditions, while language models handle unstructured interpretation; this does not directly meet the required outcome to ensure answers include only documents for a user's department and effective policy year.

C:

For a system whose objective is to keep context that crosses page or section boundaries from being lost during indexing, chunk overlap can retain nearby context across splits, although excessive overlap inflates index size and duplicates retrieval; this does not directly meet the required outcome to ensure answers include only documents for a user's department and effective policy year.

D:

For a system whose objective is to avoid fabricating an answer to a question not covered by the organization's knowledge collection, an evidence threshold and abstention path distinguish unknown facts from model completion skill; confident text is not proof of source support; this does not directly meet the required outcome to ensure answers include only documents for a user's department and effective policy year.

Question 190

A Foundry-based application must consistently authenticate a server-side generative app through Entra where supported. Which architecture decision is most defensible?

  1. Validate tool arguments before invoking actions
  2. Use token-based credentials rather than embedded keys
  3. Run independent tool requests concurrently where safe
  4. Resolve deployment and endpoint from runtime configuration

Explanation

Correct Answer

B:

For the stated requirement to authenticate a server-side generative app through Entra where supported, a workload identity and appropriate RBAC reduce static credential exposure while preserving auditable authorization decisions; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to prevent malformed model-generated parameters from reaching a write-capable API, tool schemas guide argument shape, but the application must check types, constraints, identity and authorization before execution; this does not directly meet the required outcome to authenticate a server-side generative app through Entra where supported.

C:

For a system whose objective is to reduce latency when separate read-only lookups have no ordering dependency, parallelizing independent operations shortens critical-path latency but is unsafe when one action requires another's output; this does not directly meet the required outcome to authenticate a server-side generative app through Entra where supported.

D:

For a system whose objective is to avoid routing production questions to a test model through hardcoded SDK settings, endpoint and deployment identifiers are environment-specific; runtime configuration keeps inference calls aligned to the project; this does not directly meet the required outcome to authenticate a server-side generative app through Entra where supported.

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