Question 80
A new production AI solution must detect fluent answers that introduce claims unsupported by retrieved passages. Which approach should the team select first?
- Trace each retrieval and generation span
- Monitor safety-related request outcomes
- Measure groundedness with cited source evidence
- Restrict agent outbound tool access
Explanation
Correct Answer
C:For the stated requirement to detect fluent answers that introduce claims unsupported by retrieved passages, groundedness evaluations compare generated claims with supplied evidence; fluency or user satisfaction alone cannot establish factual support; this is the mechanism that the system must implement.
Incorrect Answers
A:For a system whose objective is to locate whether a slow RAG response is delayed by search, tool calls, or model generation, distributed traces expose timing and relationships among retrieval, tool and inference spans; aggregate CPU graphs alone cannot isolate these steps; this does not directly meet the required outcome to detect fluent answers that introduce claims unsupported by retrieved passages.
B:For a system whose objective is to identify spikes in unsafe or blocked prompts and unsafe generated material, guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily; this does not directly meet the required outcome to detect fluent answers that introduce claims unsupported by retrieved passages.
D:For a system whose objective is to prevent a compromised agent from calling unapproved external endpoints, tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; this does not directly meet the required outcome to detect fluent answers that introduce claims unsupported by retrieved passages.
Question 81
An engineering team has a requirement to detect when newly approved source documents are not available to the grounding pipeline. Which capability should be in the initial design?
- Measure groundedness with cited source evidence
- Monitor retrieval freshness and ingestion errors
- Rotate compromised credentials without app downtime
- Monitor safety-related request outcomes
Explanation
Correct Answer
B:For the stated requirement to detect when newly approved source documents are not available to the grounding pipeline, indexer execution status, document counts, and source timestamps show whether retrieval data is current; a good model response cannot compensate for missing content; this is the mechanism that the system must implement.
Incorrect Answers
A:For a system whose objective is to detect fluent answers that introduce claims unsupported by retrieved passages, groundedness evaluations compare generated claims with supplied evidence; fluency or user satisfaction alone cannot establish factual support; this does not directly meet the required outcome to detect when newly approved source documents are not available to the grounding pipeline.
C:For a system whose objective is to replace an exposed secret while keeping a production inference endpoint available, credential rollover with overlapping validity and centralized secret distribution enables staged updates; longer-term managed identity can reduce secret use; this does not directly meet the required outcome to detect when newly approved source documents are not available to the grounding pipeline.
D:For a system whose objective is to identify spikes in unsafe or blocked prompts and unsafe generated material, guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily; this does not directly meet the required outcome to detect when newly approved source documents are not available to the grounding pipeline.
Question 82
A development group must locate whether a slow RAG response is delayed by search, tool calls, or model generation. What is the most appropriate Azure AI implementation?
- Trace each retrieval and generation span
- Measure groundedness with cited source evidence
- Track tool invocation errors and timeouts
- Monitor safety-related request outcomes
Explanation
Correct Answer
A:For the stated requirement to locate whether a slow RAG response is delayed by search, tool calls, or model generation, distributed traces expose timing and relationships among retrieval, tool and inference spans; aggregate CPU graphs alone cannot isolate these steps; this is the mechanism that the system must implement.
Incorrect Answers
B:For a system whose objective is to detect fluent answers that introduce claims unsupported by retrieved passages, groundedness evaluations compare generated claims with supplied evidence; fluency or user satisfaction alone cannot establish factual support; this does not directly meet the required outcome to locate whether a slow RAG response is delayed by search, tool calls, or model generation.
C:For a system whose objective is to find why agents fail to complete workflows despite healthy model endpoints, tool telemetry records HTTP failures, permission denials and timeouts at action boundaries; model token metrics will not diagnose a downstream API error; this does not directly meet the required outcome to locate whether a slow RAG response is delayed by search, tool calls, or model generation.
D:For a system whose objective is to identify spikes in unsafe or blocked prompts and unsafe generated material, guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily; this does not directly meet the required outcome to locate whether a slow RAG response is delayed by search, tool calls, or model generation.
Question 83
Before approving the proposed action to monitor safety-related request outcomes, which measurable acceptance check should be required?
- Graph cost per completed task and token usage across dialogue length bands
- Compare indexed versions and document counts with the authoritative document repository
- Track policy category, severity, false-positive review, and block rate by version
- Inspect per-stage p95 latency and correlation identifiers for slow requests
Explanation
Correct Answer
C:The accepted change is to monitor safety-related request outcomes; this evaluation checks whether it can identify spikes in unsafe or blocked prompts and unsafe generated material, since guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily.
Incorrect Answers
A:This test concerns the ability to identify waste when conversation history and tool outputs grow over successive exchanges, for which per-turn token accounting reveals input-context inflation; controlled summarization or trimming can reduce cost while preserving vital state; it does not establish whether the change to monitor safety-related request outcomes works for the current release.
B:This test concerns the ability to detect when newly approved source documents are not available to the grounding pipeline, for which indexer execution status, document counts, and source timestamps show whether retrieval data is current; a good model response cannot compensate for missing content; it does not establish whether the change to monitor safety-related request outcomes works for the current release.
D:This test concerns the ability to locate whether a slow RAG response is delayed by search, tool calls, or model generation, for which distributed traces expose timing and relationships among retrieval, tool and inference spans; aggregate CPU graphs alone cannot isolate these steps; it does not establish whether the change to monitor safety-related request outcomes works for the current release.
Question 84
A new customer-facing AI service must replace an exposed secret while keeping a production inference endpoint available. Which approach should the team select first?
- Rotate compromised credentials without app downtime
- Assign least-privilege data access to a workload identity
- Limit prompt and retrieved context length
- Restrict agent outbound tool access
Explanation
Correct Answer
A:For the stated requirement to replace an exposed secret while keeping a production inference endpoint available, credential rollover with overlapping validity and centralized secret distribution enables staged updates; longer-term managed identity can reduce secret use; this is the mechanism that the system must implement.
Incorrect Answers
B:For a system whose objective is to allow a retrieval application to query only the approved search resource without subscription administration rights, Azure RBAC can grant scoped data-plane permissions to the workload principal; broad management roles are unnecessary for query-only service calls; this does not directly meet the required outcome to replace an exposed secret while keeping a production inference endpoint available.
C:For a system whose objective is to reduce inference spend caused by oversized repeated document context, input tokens contribute to context limits and usage costs; pruning irrelevant evidence can improve efficiency without changing the model deployment; this does not directly meet the required outcome to replace an exposed secret while keeping a production inference endpoint available.
D:For a system whose objective is to prevent a compromised agent from calling unapproved external endpoints, tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; this does not directly meet the required outcome to replace an exposed secret while keeping a production inference endpoint available.
Question 85
What test should engineers run after they separate data-plane use from resource management, rather than relying solely on a successful API response?
- Attempt blocked destinations in a controlled red-team test and inspect egress logs
- Confirm no requests use the retired credential and access continues during rotation
- Test service configuration calls and document searches with each role independently
- Track 429 incidence, retry amplification, and successful completion time under load
Explanation
Correct Answer
C:The accepted change is to separate data-plane use from resource management; this evaluation checks whether it can let operators change infrastructure while limiting which identities can read sensitive index documents, since management-plane roles and data-plane roles authorize different operations; control over deployment configuration does not imply every content read permission.
Incorrect Answers
A:This test concerns the ability to prevent a compromised agent from calling unapproved external endpoints, for which tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; it does not establish whether the change to separate data-plane use from resource management works for the current release.
B:This test concerns the ability to replace an exposed secret while keeping a production inference endpoint available, for which credential rollover with overlapping validity and centralized secret distribution enables staged updates; longer-term managed identity can reduce secret use; it does not establish whether the change to separate data-plane use from resource management works for the current release.
D:This test concerns the ability to recover from temporary HTTP 429 responses without amplifying load on the model deployment, for which exponential backoff with jitter and respect for retry hints reduces synchronized retries; extra retries cannot replace insufficient sustained capacity; it does not establish whether the change to separate data-plane use from resource management works for the current release.
Question 86
A cloud engineering team must identify spikes in unsafe or blocked prompts and unsafe generated material. What is the most appropriate Azure AI implementation?
- Restrict agent outbound tool access
- Monitor safety-related request outcomes
- Monitor retrieval freshness and ingestion errors
- Trace each retrieval and generation span
Explanation
Correct Answer
B:For the stated requirement to identify spikes in unsafe or blocked prompts and unsafe generated material, guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily; this is the mechanism that the system must implement.
Incorrect Answers
A:For a system whose objective is to prevent a compromised agent from calling unapproved external endpoints, tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; this does not directly meet the required outcome to identify spikes in unsafe or blocked prompts and unsafe generated material.
C:For a system whose objective is to detect when newly approved source documents are not available to the grounding pipeline, indexer execution status, document counts, and source timestamps show whether retrieval data is current; a good model response cannot compensate for missing content; this does not directly meet the required outcome to identify spikes in unsafe or blocked prompts and unsafe generated material.
D:For a system whose objective is to locate whether a slow RAG response is delayed by search, tool calls, or model generation, distributed traces expose timing and relationships among retrieval, tool and inference spans; aggregate CPU graphs alone cannot isolate these steps; this does not directly meet the required outcome to identify spikes in unsafe or blocked prompts and unsafe generated material.
Question 87
What test should engineers run after they compare behavior across model or prompt versions, rather than relying solely on a successful API response?
- Plot tool error rates and p95 tool latency against end-to-end task completion
- Compare sustained token throughput and throttling against the new allocation
- Graph cost per completed task and token usage across dialogue length bands
- Compare task accuracy, groundedness, and safety scores by deployment revision
Explanation
Correct Answer
D:The accepted change is to compare behavior across model or prompt versions; this evaluation checks whether it can detect regression when a newly deployed model changes how it answers the same business questions, since version-tagged test results separate model changes from request-mix changes and allow a reliable rollback decision.
Incorrect Answers
A:This test concerns the ability to find why agents fail to complete workflows despite healthy model endpoints, for which tool telemetry records HTTP failures, permission denials and timeouts at action boundaries; model token metrics will not diagnose a downstream API error; it does not establish whether the change to compare behavior across model or prompt versions works for the current release.
B:This test concerns the ability to serve a stable higher request rate when the deployment repeatedly reaches its assigned token budget, for which tokens-per-minute and request limits constrain throughput; allocation or capacity decisions must match the deployment's actual measured demand; it does not establish whether the change to compare behavior across model or prompt versions works for the current release.
C:This test concerns the ability to identify waste when conversation history and tool outputs grow over successive exchanges, for which per-turn token accounting reveals input-context inflation; controlled summarization or trimming can reduce cost while preserving vital state; it does not establish whether the change to compare behavior across model or prompt versions works for the current release.
Question 88
Engineers have implemented a change to track tool invocation errors and timeouts. What validation would expose a remaining defect?
- Validate vector dimensions and semantic recall using a fixed representative query set
- Plot tool error rates and p95 tool latency against end-to-end task completion
- Attempt blocked destinations in a controlled red-team test and inspect egress logs
- Graph cost per completed task and token usage across dialogue length bands
- Track policy category, severity, false-positive review, and block rate by version
Explanation
Correct Answer
B:The accepted change is to track tool invocation errors and timeouts; this evaluation checks whether it can find why agents fail to complete workflows despite healthy model endpoints, since tool telemetry records HTTP failures, permission denials and timeouts at action boundaries; model token metrics will not diagnose a downstream API error.
Incorrect Answers
A:This test concerns the ability to catch incompatible document and query vectors when retrieval quality suddenly collapses, for which embedding dimensionality and model compatibility must be maintained between indexed passages and query vectors; vector space changes require a coordinated migration; it does not establish whether the change to track tool invocation errors and timeouts works for the current release.
C:This test concerns the ability to prevent a compromised agent from calling unapproved external endpoints, for which tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; it does not establish whether the change to track tool invocation errors and timeouts works for the current release.
D:This test concerns the ability to identify waste when conversation history and tool outputs grow over successive exchanges, for which per-turn token accounting reveals input-context inflation; controlled summarization or trimming can reduce cost while preserving vital state; it does not establish whether the change to track tool invocation errors and timeouts works for the current release.
E:This test concerns the ability to identify spikes in unsafe or blocked prompts and unsafe generated material, for which guardrail and content-filter telemetry reveals policy-related events; it should be segmented without exposing sensitive payloads unnecessarily; it does not establish whether the change to track tool invocation errors and timeouts works for the current release.
Question 89
An Azure AI workload passes basic health probes but has this issue: A contributor manages an AI Search service but a keyless document query is rejected. Which fix targets it?
- Grant the specific Search index data-reading permission at the appropriate resource scope
- Remove generic outbound reachability and allow only approved tool destinations and methods
- Assign the required data-plane reader role separately from the management role
- Instrument each tool call with duration, status, and bounded retry decisions
Explanation
Correct Answer
C:The fault signal is that a contributor manages an AI Search service but a keyless document query is rejected; the targeted fix works because management-plane roles and data-plane roles authorize different operations; control over deployment configuration does not imply every content read permission, restoring the ability to let operators change infrastructure while limiting which identities can read sensitive index documents.
Incorrect Answers
A:If the failing function were to allow a retrieval application to query only the approved search resource without subscription administration rights, this change could be relevant because Azure RBAC can grant scoped data-plane permissions to the workload principal; broad management roles are unnecessary for query-only service calls; the actual symptom is that a contributor manages an AI Search service but a keyless document query is rejected, which calls for a different repair.
B:If the failing function were to prevent a compromised agent from calling unapproved external endpoints, this change could be relevant because tool allowlists, network egress policies, and scoped identities limit what an agent can reach; prompt instructions alone are not a network control; the actual symptom is that a contributor manages an AI Search service but a keyless document query is rejected, which calls for a different repair.
D:If the failing function were to find why agents fail to complete workflows despite healthy model endpoints, this change could be relevant because tool telemetry records HTTP failures, permission denials and timeouts at action boundaries; model token metrics will not diagnose a downstream API error; the actual symptom is that a contributor manages an AI Search service but a keyless document query is rejected, which calls for a different repair.
Question 90
Engineers have implemented a change to adjust quota allocation for sustained traffic. What validation would expose a remaining defect?
- Test service configuration calls and document searches with each role independently
- Track 429 incidence, retry amplification, and successful completion time under load
- Compare sustained token throughput and throttling against the new allocation
- Measure input tokens per answered query alongside grounded-answer quality
Explanation
Correct Answer
C:The accepted change is to adjust quota allocation for sustained traffic; this evaluation checks whether it can serve a stable higher request rate when the deployment repeatedly reaches its assigned token budget, since tokens-per-minute and request limits constrain throughput; allocation or capacity decisions must match the deployment's actual measured demand.
Incorrect Answers
A:This test concerns the ability to let operators change infrastructure while limiting which identities can read sensitive index documents, for which management-plane roles and data-plane roles authorize different operations; control over deployment configuration does not imply every content read permission; it does not establish whether the change to adjust quota allocation for sustained traffic works for the current release.
B:This test concerns the ability to recover from temporary HTTP 429 responses without amplifying load on the model deployment, for which exponential backoff with jitter and respect for retry hints reduces synchronized retries; extra retries cannot replace insufficient sustained capacity; it does not establish whether the change to adjust quota allocation for sustained traffic works for the current release.
D:This test concerns the ability to reduce inference spend caused by oversized repeated document context, for which input tokens contribute to context limits and usage costs; pruning irrelevant evidence can improve efficiency without changing the model deployment; it does not establish whether the change to adjust quota allocation for sustained traffic works for the current release.