Topic 13 Practice Test 1 covers GenAI Monitoring, Model Invocation Logging, Tool Observability, and Vector Operations in Domain 4 (Task 4.3). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A support organization must capture model inputs, outputs, and invocation metadata for supported Bedrock calls. Which approach most directly satisfies the stated requirement? Choose ONE.
- end-to-end correlation ID
- guardrail intervention count
- log retention policy
- Amazon Bedrock model invocation logging
Correct Answer: D
Correct Answer
Answer D is correct because Amazon Bedrock model invocation logging is correct because it records invocation details to configured destinations for operational analysis subject to logging configuration and data-handling controls. The stem needs the team to capture model inputs, outputs, and invocation metadata for supported Bedrock calls. That fit makes Amazon Bedrock model invocation logging meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because Use end-to-end correlation ID to link application logs, tool calls, retrieval, and model invocation records for one request. Its mechanism propagates one identifier across all components so distributed evidence can be reconstructed. The current scenario requires Amazon Bedrock model invocation logging; end-to-end correlation ID solves a different design issue.
Answer B is incorrect because guardrail intervention count belongs where the governance team wants evidence about safety-policy activity and changing risk patterns. It helps to measure how often safety policies block or modify model interactions. The stem instead requires Amazon Bedrock model invocation logging, so guardrail intervention count does not meet the decisive condition.
Answer C is incorrect because log retention policy is meant to keep diagnostic records only for the required operational and compliance period. Use it when monitoring evidence is required, but privacy and cost rules limit how long detailed records may remain. This stem calls for Amazon Bedrock model invocation logging; log retention policy addresses another operational need.
Question 2
A cloud engineering team must identify who called Bedrock control-plane or supported API operations and when. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudTrail API auditing
- tool-call latency metric
- groundedness evaluation sample
- Amazon Bedrock model invocation logging
Correct Answer: A
Correct Answer
Answer A is correct because CloudTrail API auditing is correct because it records AWS API activity for audit and investigation. The stem needs the team to identify who called Bedrock control-plane or supported API operations and when. That fit makes CloudTrail API auditing meet the constraint; alternatives differ materially.
Incorrect Answers
Answer B is incorrect because The role of tool-call latency metric is to measure whether external tools, rather than the FM, dominate response time. It is useful when agent responses are slow even when model inference itself remains within target latency. Here the design needs CloudTrail API auditing; this choice instead solves an adjacent problem.
Answer C is incorrect because The role of groundedness evaluation sample is to monitor whether generated answers remain supported by retrieved evidence. It is useful when availability is healthy but the team needs a quality signal for hallucination or unsupported claims. Here the design needs CloudTrail API auditing; this choice instead solves an adjacent problem.
Answer D is incorrect because Amazon Bedrock model invocation logging belongs where the team needs inspectable inference payload evidence rather than only aggregate metrics. It helps to capture model inputs, outputs, and invocation metadata for supported Bedrock calls. The stem instead requires CloudTrail API auditing, so Amazon Bedrock model invocation logging does not meet the decisive condition.
Question 3
A infrastructure team must track tail inference latency instead of relying only on an average. Which approach most directly satisfies the stated requirement? Choose ONE.
- tool-call failure ratio
- retrieval relevance evaluation
- CloudWatch latency percentile metric
- CloudTrail API auditing
Correct Answer: C
Correct Answer
Answer C is correct because CloudWatch latency percentile metric is correct because it uses percentile statistics such as p95 to expose slow-request behavior hidden by the mean. The stem needs the team to track tail inference latency instead of relying only on an average. That fit makes CloudWatch latency percentile metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because The role of tool-call failure ratio is to detect unstable agent tools separately from model failures. It is useful when an agent frequently retries or abandons tasks because one integration is unreliable. Here the design needs CloudWatch latency percentile metric; this choice instead solves an adjacent problem.
Answer B is incorrect because retrieval relevance evaluation belongs where poor answers may result from search returning the wrong evidence even when the FM follows context correctly. It helps to measure whether the passages supplied to the model actually match the query intent. The stem instead requires CloudWatch latency percentile metric, so retrieval relevance evaluation does not meet the decisive condition.
Answer D is incorrect because The role of CloudTrail API auditing is to identify who called Bedrock control-plane or supported API operations and when. It is useful when the requirement is account-level API attribution rather than model-response quality monitoring. Here the design needs CloudWatch latency percentile metric; this choice instead solves an adjacent problem.
Question 4
A architecture review board must measure how often requests are rejected or delayed by service limits. Which approach most directly satisfies the stated requirement? Choose ONE.
- vector retrieval latency metric
- CloudWatch throttling-rate metric
- SLO alarm on sustained breach
- CloudWatch latency percentile metric
Correct Answer: B
Correct Answer
Answer B is correct because CloudWatch throttling-rate metric is correct because it tracks throttling errors relative to request volume so capacity pressure is visible. The stem needs the team to measure how often requests are rejected or delayed by service limits. That fit makes CloudWatch throttling-rate metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because Use vector retrieval latency metric to observe the search component of a RAG request independently. Its mechanism measures query-to-results time at the vector store so retrieval bottlenecks are visible. The current scenario requires CloudWatch throttling-rate metric; vector retrieval latency metric solves a different design issue.
Answer C is incorrect because SLO alarm on sustained breach is meant to page operators only when a service objective is violated for a meaningful interval. Use it when the team needs actionable alerting for latency or error objectives rather than dashboard-only visibility. This stem calls for CloudWatch throttling-rate metric; SLO alarm on sustained breach addresses another operational need.
Answer D is incorrect because CloudWatch latency percentile metric is meant to track tail inference latency instead of relying only on an average. Use it when most calls are fast but a small slow tail violates the user experience objective. This stem calls for CloudWatch throttling-rate metric; CloudWatch latency percentile metric addresses another operational need.
Question 5
A production operations group must distinguish service-side failures from application validation errors. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudWatch server-error rate
- vector index freshness metric
- service dashboard by model and workflow
- CloudWatch throttling-rate metric
Correct Answer: A
Correct Answer
Answer A is correct because CloudWatch server-error rate is correct because it monitors 5xx-style service failures separately from client-side request issues. The stem needs the team to distinguish service-side failures from application validation errors. That fit makes CloudWatch server-error rate meet the constraint; alternatives differ materially.
Incorrect Answers
Answer B is incorrect because The role of vector index freshness metric is to verify source changes become searchable within the required interval. It is useful when users report answers based on documents that were updated recently but have not appeared in retrieval yet. Here the design needs CloudWatch server-error rate; this choice instead solves an adjacent problem.
Answer C is incorrect because service dashboard by model and workflow can be appropriate because it groups latency, errors, throttles, usage, retrieval, and tool metrics into operational views. Apply it where operators need rapid isolation of which model or workflow is degrading. This case depends on CloudWatch server-error rate, leaving service dashboard by model and workflow mismatched to the requirement.
Answer D is incorrect because The role of CloudWatch throttling-rate metric is to measure how often requests are rejected or delayed by service limits. It is useful when users see intermittent rate-limit failures during traffic peaks. Here the design needs CloudWatch server-error rate; this choice instead solves an adjacent problem.
Question 6
A platform team must measure prompt and completion token consumption over time. Which approach most directly satisfies the stated requirement? Choose ONE.
- ingestion failure counter
- sensitive-field redaction before log storage
- token usage metric
- CloudWatch server-error rate
Correct Answer: C
Correct Answer
Answer C is correct because token usage metric is correct because it records request token counts so cost and context growth can be analyzed. The stem needs the team to measure prompt and completion token consumption over time. That fit makes token usage metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because ingestion failure counter can be appropriate because it counts and classifies failed ingestion events for repair or replay. Apply it where some source files silently disappear from the searchable knowledge base. This case depends on token usage metric, leaving ingestion failure counter mismatched to the requirement.
Answer B is incorrect because sensitive-field redaction before log storage can be appropriate because it removes or masks sensitive fields before durable logging while preserving diagnostic metadata. Apply it where invocation or tool payloads may contain credentials, PII, or confidential business data. This case depends on token usage metric, leaving sensitive-field redaction before log storage mismatched to the requirement.
Answer D is incorrect because CloudWatch server-error rate works by monitors 5xx-style service failures separately from client-side request issues. It fits where the operations team needs an SLO signal for provider-side inference reliability. The required function here is token usage metric, so this option targets the wrong condition.
Question 7
A network operations group must separate spend by model, application, or tenant tags and usage records. Which approach most directly satisfies the stated requirement? Choose ONE.
- guardrail intervention count
- per-model cost attribution
- log retention policy
- token usage metric
Correct Answer: B
Correct Answer
Answer B is correct because per-model cost attribution is correct because it combines usage dimensions with application metadata so expensive workloads are identifiable. The stem needs the team to separate spend by model, application, or tenant tags and usage records. That fit makes per-model cost attribution meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because guardrail intervention count belongs where the governance team wants evidence about safety-policy activity and changing risk patterns. It helps to measure how often safety policies block or modify model interactions. The stem instead requires per-model cost attribution, so guardrail intervention count does not meet the decisive condition.
Answer C is incorrect because log retention policy is meant to keep diagnostic records only for the required operational and compliance period. Use it when monitoring evidence is required, but privacy and cost rules limit how long detailed records may remain. This stem calls for per-model cost attribution; log retention policy addresses another operational need.
Answer D is incorrect because token usage metric is meant to measure prompt and completion token consumption over time. Use it when spend is increasing and the team needs evidence about which prompts or responses consume the budget. This stem calls for per-model cost attribution; token usage metric addresses another operational need.
Question 8
A application team must link application logs, tool calls, retrieval, and model invocation records for one request. Which approach most directly satisfies the stated requirement? Choose ONE.
- groundedness evaluation sample
- Amazon Bedrock model invocation logging
- per-model cost attribution
- end-to-end correlation ID
Correct Answer: D
Correct Answer
Answer D is correct because end-to-end correlation ID is correct because it propagates one identifier across all components so distributed evidence can be reconstructed. The stem needs the team to link application logs, tool calls, retrieval, and model invocation records for one request. That fit makes end-to-end correlation ID meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because The role of groundedness evaluation sample is to monitor whether generated answers remain supported by retrieved evidence. It is useful when availability is healthy but the team needs a quality signal for hallucination or unsupported claims. Here the design needs end-to-end correlation ID; this choice instead solves an adjacent problem.
Answer B is incorrect because Amazon Bedrock model invocation logging belongs where the team needs inspectable inference payload evidence rather than only aggregate metrics. It helps to capture model inputs, outputs, and invocation metadata for supported Bedrock calls. The stem instead requires end-to-end correlation ID, so Amazon Bedrock model invocation logging does not meet the decisive condition.
Answer C is incorrect because Use per-model cost attribution to separate spend by model, application, or tenant tags and usage records. Its mechanism combines usage dimensions with application metadata so expensive workloads are identifiable. The current scenario requires end-to-end correlation ID; per-model cost attribution solves a different design issue.
Question 9
A enterprise architect must measure whether external tools, rather than the FM, dominate response time. Which approach most directly satisfies the stated requirement? Choose ONE.
- retrieval relevance evaluation
- tool-call latency metric
- CloudTrail API auditing
- end-to-end correlation ID
Correct Answer: B
Correct Answer
Answer B is correct because tool-call latency metric is correct because it records start/end timing for each tool invocation and aggregates by tool. The stem needs the team to measure whether external tools, rather than the FM, dominate response time. That fit makes tool-call latency metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because retrieval relevance evaluation belongs where poor answers may result from search returning the wrong evidence even when the FM follows context correctly. It helps to measure whether the passages supplied to the model actually match the query intent. The stem instead requires tool-call latency metric, so retrieval relevance evaluation does not meet the decisive condition.
Answer C is incorrect because The role of CloudTrail API auditing is to identify who called Bedrock control-plane or supported API operations and when. It is useful when the requirement is account-level API attribution rather than model-response quality monitoring. Here the design needs tool-call latency metric; this choice instead solves an adjacent problem.
Answer D is incorrect because end-to-end correlation ID is meant to link application logs, tool calls, retrieval, and model invocation records for one request. Use it when a single answer passes through retrieval and tools and failures cannot otherwise be joined reliably. This stem calls for tool-call latency metric; end-to-end correlation ID addresses another operational need.
Question 10
A site reliability team must detect unstable agent tools separately from model failures. Which approach most directly satisfies the stated requirement? Choose ONE.
- SLO alarm on sustained breach
- CloudWatch latency percentile metric
- tool-call latency metric
- tool-call failure ratio
Correct Answer: D
Correct Answer
Answer D is correct because tool-call failure ratio is correct because it tracks successful and failed tool invocations by tool and error class. The stem needs the team to detect unstable agent tools separately from model failures. That fit makes tool-call failure ratio meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because SLO alarm on sustained breach is meant to page operators only when a service objective is violated for a meaningful interval. Use it when the team needs actionable alerting for latency or error objectives rather than dashboard-only visibility. This stem calls for tool-call failure ratio; SLO alarm on sustained breach addresses another operational need.
Answer B is incorrect because CloudWatch latency percentile metric is meant to track tail inference latency instead of relying only on an average. Use it when most calls are fast but a small slow tail violates the user experience objective. This stem calls for tool-call failure ratio; CloudWatch latency percentile metric addresses another operational need.
Answer C is incorrect because Use tool-call latency metric to measure whether external tools, rather than the FM, dominate response time. Its mechanism records start/end timing for each tool invocation and aggregates by tool. The current scenario requires tool-call failure ratio; tool-call latency metric solves a different design issue.
Question 11
A security engineering group must observe the search component of a RAG request independently. Which approach most directly satisfies the stated requirement? Choose ONE.
- vector retrieval latency metric
- service dashboard by model and workflow
- CloudWatch throttling-rate metric
- tool-call failure ratio
Correct Answer: A
Correct Answer
Answer A is correct because vector retrieval latency metric is correct because it measures query-to-results time at the vector store so retrieval bottlenecks are visible. The stem needs the team to observe the search component of a RAG request independently. That fit makes vector retrieval latency metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer B is incorrect because service dashboard by model and workflow can be appropriate because it groups latency, errors, throttles, usage, retrieval, and tool metrics into operational views. Apply it where operators need rapid isolation of which model or workflow is degrading. This case depends on vector retrieval latency metric, leaving service dashboard by model and workflow mismatched to the requirement.
Answer C is incorrect because The role of CloudWatch throttling-rate metric is to measure how often requests are rejected or delayed by service limits. It is useful when users see intermittent rate-limit failures during traffic peaks. Here the design needs vector retrieval latency metric; this choice instead solves an adjacent problem.
Answer D is incorrect because Use tool-call failure ratio to detect unstable agent tools separately from model failures. Its mechanism tracks successful and failed tool invocations by tool and error class. The current scenario requires vector retrieval latency metric; tool-call failure ratio solves a different design issue.
Question 12
A branch deployment team must verify source changes become searchable within the required interval. Which approach most directly satisfies the stated requirement? Choose ONE.
- sensitive-field redaction before log storage
- CloudWatch server-error rate
- vector index freshness metric
- vector retrieval latency metric
Correct Answer: C
Correct Answer
Answer C is correct because vector index freshness metric is correct because it measures lag between source modification and successful vector-index update. The stem needs the team to verify source changes become searchable within the required interval. That fit makes vector index freshness metric meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because sensitive-field redaction before log storage can be appropriate because it removes or masks sensitive fields before durable logging while preserving diagnostic metadata. Apply it where invocation or tool payloads may contain credentials, PII, or confidential business data. This case depends on vector index freshness metric, leaving sensitive-field redaction before log storage mismatched to the requirement.
Answer B is incorrect because CloudWatch server-error rate works by monitors 5xx-style service failures separately from client-side request issues. It fits where the operations team needs an SLO signal for provider-side inference reliability. The required function here is vector index freshness metric, so this option targets the wrong condition.
Answer D is incorrect because vector retrieval latency metric is meant to observe the search component of a RAG request independently. Use it when RAG response time increased after an index or corpus change while model latency stayed stable. This stem calls for vector index freshness metric; vector retrieval latency metric addresses another operational need.
Question 13
A support organization must surface documents or chunks that failed parsing, embedding, or indexing. Which approach most directly satisfies the stated requirement? Choose ONE.
- ingestion failure counter
- log retention policy
- token usage metric
- vector index freshness metric
Correct Answer: A
Correct Answer
Answer A is correct because ingestion failure counter is correct because it counts and classifies failed ingestion events for repair or replay. The stem needs the team to surface documents or chunks that failed parsing, embedding, or indexing. That fit makes ingestion failure counter meet the constraint; alternatives differ materially.
Incorrect Answers
Answer B is incorrect because log retention policy is meant to keep diagnostic records only for the required operational and compliance period. Use it when monitoring evidence is required, but privacy and cost rules limit how long detailed records may remain. This stem calls for ingestion failure counter; log retention policy addresses another operational need.
Answer C is incorrect because token usage metric is meant to measure prompt and completion token consumption over time. Use it when spend is increasing and the team needs evidence about which prompts or responses consume the budget. This stem calls for ingestion failure counter; token usage metric addresses another operational need.
Answer D is incorrect because Use vector index freshness metric to verify source changes become searchable within the required interval. Its mechanism measures lag between source modification and successful vector-index update. The current scenario requires ingestion failure counter; vector index freshness metric solves a different design issue.
Question 14
A cloud engineering team must measure how often safety policies block or modify model interactions. Which approach most directly satisfies the stated requirement? Choose ONE.
- Amazon Bedrock model invocation logging
- per-model cost attribution
- ingestion failure counter
- guardrail intervention count
Correct Answer: D
Correct Answer
Answer D is correct because guardrail intervention count is correct because it tracks guardrail evaluations and outcomes without treating them as generic model errors. The stem needs the team to measure how often safety policies block or modify model interactions. That fit makes guardrail intervention count meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because Amazon Bedrock model invocation logging belongs where the team needs inspectable inference payload evidence rather than only aggregate metrics. It helps to capture model inputs, outputs, and invocation metadata for supported Bedrock calls. The stem instead requires guardrail intervention count, so Amazon Bedrock model invocation logging does not meet the decisive condition.
Answer B is incorrect because Use per-model cost attribution to separate spend by model, application, or tenant tags and usage records. Its mechanism combines usage dimensions with application metadata so expensive workloads are identifiable. The current scenario requires guardrail intervention count; per-model cost attribution solves a different design issue.
Answer C is incorrect because ingestion failure counter belongs where some source files silently disappear from the searchable knowledge base. It helps to surface documents or chunks that failed parsing, embedding, or indexing. The stem instead requires guardrail intervention count, so ingestion failure counter does not meet the decisive condition.
Question 15
A infrastructure team must monitor whether generated answers remain supported by retrieved evidence. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudTrail API auditing
- end-to-end correlation ID
- groundedness evaluation sample
- guardrail intervention count
Correct Answer: C
Correct Answer
Answer C is correct because groundedness evaluation sample is correct because it periodically scores sampled responses for factual support against supplied context. The stem needs the team to monitor whether generated answers remain supported by retrieved evidence. That fit makes groundedness evaluation sample meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because The role of CloudTrail API auditing is to identify who called Bedrock control-plane or supported API operations and when. It is useful when the requirement is account-level API attribution rather than model-response quality monitoring. Here the design needs groundedness evaluation sample; this choice instead solves an adjacent problem.
Answer B is incorrect because end-to-end correlation ID is meant to link application logs, tool calls, retrieval, and model invocation records for one request. Use it when a single answer passes through retrieval and tools and failures cannot otherwise be joined reliably. This stem calls for groundedness evaluation sample; end-to-end correlation ID addresses another operational need.
Answer D is incorrect because guardrail intervention count works by tracks guardrail evaluations and outcomes without treating them as generic model errors. It fits where the governance team wants evidence about safety-policy activity and changing risk patterns. The required function here is groundedness evaluation sample, so this option targets the wrong condition.
Question 16
A architecture review board must measure whether the passages supplied to the model actually match the query intent. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudWatch latency percentile metric
- retrieval relevance evaluation
- tool-call latency metric
- groundedness evaluation sample
Correct Answer: B
Correct Answer
Answer B is correct because retrieval relevance evaluation is correct because it scores retrieved candidates independently from generation quality. The stem needs the team to measure whether the passages supplied to the model actually match the query intent. That fit makes retrieval relevance evaluation meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because CloudWatch latency percentile metric is meant to track tail inference latency instead of relying only on an average. Use it when most calls are fast but a small slow tail violates the user experience objective. This stem calls for retrieval relevance evaluation; CloudWatch latency percentile metric addresses another operational need.
Answer C is incorrect because Use tool-call latency metric to measure whether external tools, rather than the FM, dominate response time. Its mechanism records start/end timing for each tool invocation and aggregates by tool. The current scenario requires retrieval relevance evaluation; tool-call latency metric solves a different design issue.
Answer D is incorrect because Use groundedness evaluation sample to monitor whether generated answers remain supported by retrieved evidence. Its mechanism periodically scores sampled responses for factual support against supplied context. The current scenario requires retrieval relevance evaluation; groundedness evaluation sample solves a different design issue.
Question 17
A production operations group must page operators only when a service objective is violated for a meaningful interval. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudWatch throttling-rate metric
- SLO alarm on sustained breach
- tool-call failure ratio
- retrieval relevance evaluation
Correct Answer: B
Correct Answer
Answer B is correct because SLO alarm on sustained breach is correct because it combines threshold and duration so transient noise does not create unnecessary incidents. The stem needs the team to page operators only when a service objective is violated for a meaningful interval. That fit makes SLO alarm on sustained breach meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because The role of CloudWatch throttling-rate metric is to measure how often requests are rejected or delayed by service limits. It is useful when users see intermittent rate-limit failures during traffic peaks. Here the design needs SLO alarm on sustained breach; this choice instead solves an adjacent problem.
Answer C is incorrect because Use tool-call failure ratio to detect unstable agent tools separately from model failures. Its mechanism tracks successful and failed tool invocations by tool and error class. The current scenario requires SLO alarm on sustained breach; tool-call failure ratio solves a different design issue.
Answer D is incorrect because retrieval relevance evaluation works by scores retrieved candidates independently from generation quality. It fits where poor answers may result from search returning the wrong evidence even when the FM follows context correctly. The required function here is SLO alarm on sustained breach, so this option targets the wrong condition.
Question 18
A platform team must compare health signals across models, applications, and processing stages. Which approach most directly satisfies the stated requirement? Choose ONE.
- CloudWatch server-error rate
- vector retrieval latency metric
- service dashboard by model and workflow
- SLO alarm on sustained breach
Correct Answer: C
Correct Answer
Answer C is correct because service dashboard by model and workflow is correct because it groups latency, errors, throttles, usage, retrieval, and tool metrics into operational views. The stem needs the team to compare health signals across models, applications, and processing stages. That fit makes service dashboard by model and workflow meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because CloudWatch server-error rate works by monitors 5xx-style service failures separately from client-side request issues. It fits where the operations team needs an SLO signal for provider-side inference reliability. The required function here is service dashboard by model and workflow, so this option targets the wrong condition.
Answer B is incorrect because vector retrieval latency metric is meant to observe the search component of a RAG request independently. Use it when RAG response time increased after an index or corpus change while model latency stayed stable. This stem calls for service dashboard by model and workflow; vector retrieval latency metric addresses another operational need.
Answer D is incorrect because The role of SLO alarm on sustained breach is to page operators only when a service objective is violated for a meaningful interval. It is useful when the team needs actionable alerting for latency or error objectives rather than dashboard-only visibility. Here the design needs service dashboard by model and workflow; this choice instead solves an adjacent problem.
Question 19
A network operations group must keep observability useful without copying secrets or regulated data into logs. Which approach most directly satisfies the stated requirement? Choose ONE.
- token usage metric
- vector index freshness metric
- service dashboard by model and workflow
- sensitive-field redaction before log storage
Correct Answer: D
Correct Answer
Answer D is correct because sensitive-field redaction before log storage is correct because it removes or masks sensitive fields before durable logging while preserving diagnostic metadata. The stem needs the team to keep observability useful without copying secrets or regulated data into logs. That fit makes sensitive-field redaction before log storage meet the constraint; alternatives differ materially.
Incorrect Answers
Answer A is incorrect because token usage metric is meant to measure prompt and completion token consumption over time. Use it when spend is increasing and the team needs evidence about which prompts or responses consume the budget. This stem calls for sensitive-field redaction before log storage; token usage metric addresses another operational need.
Answer B is incorrect because Use vector index freshness metric to verify source changes become searchable within the required interval. Its mechanism measures lag between source modification and successful vector-index update. The current scenario requires sensitive-field redaction before log storage; vector index freshness metric solves a different design issue.
Answer C is incorrect because service dashboard by model and workflow belongs where operators need rapid isolation of which model or workflow is degrading. It helps to compare health signals across models, applications, and processing stages. The stem instead requires sensitive-field redaction before log storage, so service dashboard by model and workflow does not meet the decisive condition.
Question 20
A application team must keep diagnostic records only for the required operational and compliance period. Which approach most directly satisfies the stated requirement? Choose ONE.
- log retention policy
- per-model cost attribution
- ingestion failure counter
- sensitive-field redaction before log storage
Correct Answer: A
Correct Answer
Answer A is correct because log retention policy is correct because it applies explicit retention to log destinations so data does not persist indefinitely. The stem needs the team to keep diagnostic records only for the required operational and compliance period. That fit makes log retention policy meet the constraint; alternatives differ materially.
Incorrect Answers
Answer B is incorrect because Use per-model cost attribution to separate spend by model, application, or tenant tags and usage records. Its mechanism combines usage dimensions with application metadata so expensive workloads are identifiable. The current scenario requires log retention policy; per-model cost attribution solves a different design issue.
Answer C is incorrect because ingestion failure counter belongs where some source files silently disappear from the searchable knowledge base. It helps to surface documents or chunks that failed parsing, embedding, or indexing. The stem instead requires log retention policy, so ingestion failure counter does not meet the decisive condition.
Answer D is incorrect because sensitive-field redaction before log storage belongs where invocation or tool payloads may contain credentials, PII, or confidential business data. It helps to keep observability useful without copying secrets or regulated data into logs. The stem instead requires log retention policy, so sensitive-field redaction before log storage does not meet the decisive condition.
