Topic 15 Practice Test 2 covers GenAI Troubleshooting, Trace Analysis, Failure Isolation, and Remediation in Domain 5 (Task 5.2). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
The API returns throttling errors during traffic spikes, while the same prompt succeeds when retried later. What should the team implement? Choose ONE.
- Rate-limit/backoff handling
- Structured API error logging
- Circuit breaker around failing tool
- Timeout isolation
Correct Answer: A
Correct Answer
Answer A is correct because Rate-limit/backoff handling — in this scenario, recognizes throttling and retries according, in this case, to bounded exponential backoff and, in this case, service guidance. Useful when requests fail intermittently with throttling, at this layer, rather than content errors. Need: handle service throttling as a transient integration condition. Rate-limit/backoff handling matches it directly.
Incorrect Answers
Answer B is incorrect because Structured API error logging — in the current architecture, captures request identifiers, status codes,, in the described flow, exception details, and relevant non-sensitive, in the described flow, metadata for failed calls. Applies where integration failures must be diagnosed, in this case, from concrete API evidence. Target outcome: handle service throttling as a transient integration condition. Structured API error logging misses that outcome.
Answer C is incorrect because A circuit breaker protects an application from repeatedly calling a persistently failing downstream dependency after an error threshold is reached. The described failures are transient service throttles that succeed later, so bounded retry behavior with exponential backoff is the appropriate response. A circuit breaker addresses a different failure mode.
Answer D is incorrect because Timeout isolation — for this design, separates model, retrieval, and tool, for this workload, timing so the component exceeding, for this workload, its time budget can be, for this workload, identified. Fits when the overall request times out, for the stated need, but the slow stage is, for the stated need, unknown. Constraint: handle service throttling as a transient integration condition. Timeout isolation targets another layer.
Question 2
Users see a 30-second timeout, but the team cannot tell whether retrieval, model inference, or a tool call consumed the budget. Which option most directly addresses the requirement? Choose ONE.
- AWS X-Ray tracing
- Response-shape analysis
- Timeout isolation
- CloudWatch Logs correlation
Correct Answer: C
Correct Answer
Answer C is correct because Timeout isolation — at this decision point, separates model, retrieval, and tool, within this design, timing so the component exceeding, within this design, its time budget can be, within this design, identified. Typical use: the overall request times out, in the described flow, but the slow stage is, in the described flow, unknown. Goal: identify which request stage exceeds its time budget. Timeout isolation provides that behavior.
Incorrect Answers
Answer A is incorrect because AWS X-Ray tracing — under these constraints, traces distributed request paths and, at this layer, timing so latency or failures, at this layer, can be isolated across application, at this layer, components. Choose it when a multi-service GenAI request has, under these constraints, an unclear latency or failure, under these constraints, bottleneck. Requirement: identify which request stage exceeds its time budget. AWS X-Ray tracing is not the direct fit.
Answer B is incorrect because Response-shape analysis — under these constraints, examines returned status, metadata, stop, at this layer, reason, and payload structure to, at this layer, distinguish model behavior from integration, at this layer, parsing errors. Choose it when the caller receives a response, under these constraints, but handles it incorrectly. Requirement: identify which request stage exceeds its time budget. Response-shape analysis is not the direct fit.
Answer D is incorrect because CloudWatch Logs correlation — in the current architecture, centralizes application and service logs, in the described flow, so request IDs and events, in the described flow, can be correlated across failed, in the described flow, interactions. Applies where troubleshooting requires a timeline of, in this case, logged events. Target outcome: identify which request stage exceeds its time budget. CloudWatch Logs correlation misses that outcome.
Question 3
Relevant documents exist in the index but a metadata filter excludes the correct product version and top-k is set too low. Which design best satisfies this requirement? Choose ONE.
- Source-document freshness check
- Top-k and metadata-filter review
- Retrieval relevance analysis
- Vector-search performance tuning
Correct Answer: B
Correct Answer
Answer B is correct because Top-k and metadata-filter review — under these constraints, checks retrieval count and filters, at this layer, to confirm that relevant documents, at this layer, are not excluded or drowned, at this layer, out. Choose it when the index is healthy but, under these constraints, query configuration suppresses useful context. Requirement: fix query configuration that suppresses useful context. Top-k and metadata-filter review addresses it precisely.
Incorrect Answers
Answer A is incorrect because Source-document freshness check — under these constraints, confirms the retrieval corpus contains, at this layer, the current source content and, at this layer, that updates were indexed. Choose it when answers use outdated facts because, under these constraints, the knowledge source may be, under these constraints, stale. Requirement: fix query configuration that suppresses useful context. Source-document freshness check is not the direct fit.
Answer C is incorrect because Retrieval relevance analysis — at this decision point, checks whether retrieved chunks actually, within this design, match the question and contain, within this design, the evidence the model needs. Typical use: RAG answers are wrong because, in the described flow, supporting context may be irrelevant. Goal: fix query configuration that suppresses useful context. Retrieval relevance analysis does not provide it.
Answer D is incorrect because Vector-search performance tuning — at this decision point, optimizes vector index/search parameters and, within this design, query patterns to reduce latency, within this design, while preserving acceptable recall. Typical use: retrieval is relevant but too slow. Goal: fix query configuration that suppresses useful context. Vector-search performance tuning does not provide it.
Question 4
The knowledge source was updated yesterday, but answers still quote last quarter’s policy. What is the most appropriate implementation choice? Choose ONE.
- Chunking/preprocessing remediation
- Embedding drift monitoring
- Retrieval relevance analysis
- Source-document freshness check
Correct Answer: D
Correct Answer
Answer D is correct because Source-document freshness check — for the described workload, confirms the retrieval corpus contains, for the stated need, the current source content and, for the stated need, that updates were indexed. Best suited when answers use outdated facts because, for this decision, the knowledge source may be, for this decision, stale. Deciding factor: verify that current source content was ingested and indexed. Source-document freshness check fits this case.
Incorrect Answers
Answer A is incorrect because Chunking/preprocessing remediation — under these constraints, fixes document splitting, normalization, metadata,, at this layer, or preprocessing that prevents useful, at this layer, passages from being retrieved. Choose it when source text is present but, under these constraints, indexing preparation damages retrieval. Requirement: verify that current source content was ingested and indexed. Chunking/preprocessing remediation is not the direct fit.
Answer B is incorrect because Embedding drift monitoring — for the described workload, detects changes in retrieval behavior, for the stated need, or vector distributions after model,, for the stated need, corpus, or preprocessing changes. Best suited when retrieval quality degrades gradually after, for this decision, upstream changes. Deciding factor: verify that current source content was ingested and indexed. Embedding drift monitoring addresses something else.
Answer C is incorrect because Retrieval relevance analysis — under these constraints, checks whether retrieved chunks actually, at this layer, match the question and contain, at this layer, the evidence the model needs. Choose it when RAG answers are wrong because, under these constraints, supporting context may be irrelevant. Requirement: verify that current source content was ingested and indexed. Retrieval relevance analysis is not the direct fit.
Question 5
The model streams all chunks, but the client library drops the final event when assembling the response. Which approach should the developer use? Choose ONE.
- Response-shape analysis
- Truncation-error analysis
- Parser/serialization check
- Structured API error logging
Correct Answer: C
Correct Answer
Answer C is correct because Parser/serialization check — at this decision point, verifies that the application correctly, within this design, parses streamed or structured model, within this design, responses without dropping fields or, within this design, chunks. Typical use: the service returns data but, in the described flow, the application presents incomplete output. Goal: debug application handling of a valid streamed response. Parser/serialization check provides that behavior.
Incorrect Answers
Answer A is incorrect because Response-shape analysis — for the described workload, examines returned status, metadata, stop, for the stated need, reason, and payload structure to, for the stated need, distinguish model behavior from integration, for the stated need, parsing errors. Best suited when the caller receives a response, for this decision, but handles it incorrectly. Deciding factor: debug application handling of a valid streamed response. Response-shape analysis addresses something else.
Answer B is incorrect because Truncation-error analysis — at this decision point, inspects where input or output, within this design, was cut off and correlates, within this design, it with token or payload, within this design, limits. Typical use: the symptom is incomplete content, in the described flow, at a boundary. Goal: debug application handling of a valid streamed response. Truncation-error analysis does not provide it.
Answer D is incorrect because Structured API error logging — for this operational need, captures request identifiers, status codes,, under these constraints, exception details, and relevant non-sensitive, under these constraints, metadata for failed calls. Architects use it when integration failures must be diagnosed, for this workload, from concrete API evidence. Key need: debug application handling of a valid streamed response. Structured API error logging does not satisfy it.
Question 6
An agent chooses a search tool correctly but passes an unsupported region value and receives a tool error. What should the team implement? Choose ONE.
- Request-schema validation
- CloudWatch Logs correlation
- Tool-call trace analysis
- Lambda parameter validation
Correct Answer: C
Correct Answer
Answer C is correct because Tool-call trace analysis — for this design, inspects agent/tool traces to determine, for this workload, whether the model chose the, for this workload, correct tool, supplied valid parameters,, for this workload, and received a usable result. Fits when an agent fails during multi-step, for the stated need, tool execution. Constraint: inspect the agent’s tool selection, parameters, and returned error. Tool-call trace analysis is the closest fit.
Incorrect Answers
Answer A is incorrect because Request-schema validation — for the described workload, checks required fields, types, ranges,, for the stated need, and payload structure before sending, for the stated need, an FM request. Best suited when invalid or malformed requests are, for this decision, reaching the model API. Deciding factor: inspect the agent’s tool selection, parameters, and returned error. Request-schema validation addresses something else.
Answer B is incorrect because CloudWatch Logs correlation — for this operational need, centralizes application and service logs, under these constraints, so request IDs and events, under these constraints, can be correlated across failed, under these constraints, interactions. Architects use it when troubleshooting requires a timeline of, for this workload, logged events. Key need: inspect the agent’s tool selection, parameters, and returned error. CloudWatch Logs correlation does not satisfy it.
Answer D is incorrect because Lambda parameter validation — at this decision point, validates and normalizes tool inputs, within this design, before business logic or downstream, within this design, API execution. Use it where tool calls need defensive checks, in the described flow, before side effects occur. The design requires inspect the agent’s tool selection, parameters, and returned error. Lambda parameter validation is not the direct choice.
Question 7
Only prompts containing certain categories fail, and logs show a safety intervention rather than a model service error. Which option most directly addresses the requirement? Choose ONE.
- Prompt design optimization
- Request-schema validation
- Structured API error logging
- Guardrail-intervention analysis
Correct Answer: D
Correct Answer
Answer D is correct because Guardrail-intervention analysis — in this scenario, checks safety/guardrail outcomes to distinguish, in this case, blocked content from model or, in this case, application failures. Useful when requests appear to fail only, at this layer, for certain content classes. Need: distinguish a safety-policy block from an integration failure. Guardrail-intervention analysis matches it directly.
Incorrect Answers
Answer A is incorrect because Prompt design optimization — for the described workload, restructures instructions, context ordering, delimiters,, for the stated need, or examples to make model, for the stated need, behavior more reliable. Best suited when the model receives all necessary, for this decision, data but interprets the prompt, for this decision, poorly. Deciding factor: distinguish a safety-policy block from an integration failure. Prompt design optimization addresses something else.
Answer B is incorrect because Request-schema validation — in the current architecture, checks required fields, types, ranges,, in the described flow, and payload structure before sending, in the described flow, an FM request. Applies where invalid or malformed requests are, in this case, reaching the model API. Target outcome: distinguish a safety-policy block from an integration failure. Request-schema validation misses that outcome.
Answer C is incorrect because Structured API error logging — within this implementation, captures request identifiers, status codes,, for this decision, exception details, and relevant non-sensitive, for this decision, metadata for failed calls. Strong fit when integration failures must be diagnosed, within this design, from concrete API evidence. Operational need: distinguish a safety-policy block from an integration failure. Structured API error logging is secondary here.
Question 8
Quality dropped after changing temperature, model version, and max tokens together. The team needs a known-good baseline to isolate the cause. Which design best satisfies this requirement? Choose ONE.
- Prompt test harness
- Systematic prompt refinement
- Prompt-version comparison
- Model/configuration comparison
Correct Answer: D
Correct Answer
Answer D is correct because Model/configuration comparison — for this design, reproduces the same case with, for this workload, a known-good model configuration to, for this workload, isolate whether parameters or model, for this workload, choice cause the symptom. Fits when quality changed after model or, for the stated need, inference configuration changes. Constraint: compare against a known-good model/inference configuration. Model/configuration comparison is the closest fit.
Incorrect Answers
Answer A is incorrect because Prompt test harness — in the current architecture, runs controlled prompt cases repeatedly, in the described flow, so changes can be compared, in the described flow, against expected behaviors. Applies where prompt quality regressions need reproducible, in this case, testing rather than ad hoc, in this case, editing. Target outcome: compare against a known-good model/inference configuration. Prompt test harness misses that outcome.
Answer B is incorrect because Systematic prompt refinement — under these constraints, changes one prompt factor at, at this layer, a time using measured results, at this layer, to isolate the cause of, at this layer, poor responses. Choose it when the prompt is suspected but, under these constraints, the failure source is not, under these constraints, yet isolated. Requirement: compare against a known-good model/inference configuration. Systematic prompt refinement is not the direct fit.
Answer C is incorrect because Prompt-version comparison — for the described workload, compares outputs and metrics from, for the stated need, two prompt versions against the, for the stated need, same evaluation cases. Best suited when a recent prompt change is, for this decision, suspected of causing quality degradation. Deciding factor: compare against a known-good model/inference configuration. Prompt-version comparison addresses something else.
Question 9
Long conversations fail only after many turns, while short conversations are reliable. What is the most appropriate implementation choice? Choose ONE.
- Context-window overflow diagnostics
- Truncation-error analysis
- Prompt design optimization
- Dynamic chunking
Correct Answer: A
Correct Answer
Answer A is correct because Inspect accumulated token usage, prompt growth, truncation behavior, and the model’s context limit. Reliability degrading only after many conversation turns strongly indicates that required information may be falling outside the usable context. Context-window overflow diagnostics therefore isolate the likely boundary condition before the team changes chunking or prompt design.
Incorrect Answers
Answer B is incorrect because Truncation-error analysis — under these constraints, inspects where input or output, at this layer, was cut off and correlates, at this layer, it with token or payload, at this layer, limits. Choose it when the symptom is incomplete content, under these constraints, at a boundary. Requirement: check accumulated token usage and context limits. Truncation-error analysis is not the direct fit.
Answer C is incorrect because Prompt design optimization — in the current architecture, restructures instructions, context ordering, delimiters,, in the described flow, or examples to make model, in the described flow, behavior more reliable. Applies where the model receives all necessary, in this case, data but interprets the prompt, in this case, poorly. Target outcome: check accumulated token usage and context limits. Prompt design optimization misses that outcome.
Answer D is incorrect because Dynamic chunking — under these constraints, adjusts document or context chunk, at this layer, sizes so relevant content fits, at this layer, model and retrieval constraints more, at this layer, effectively. Choose it when fixed chunks are causing important, under these constraints, passages to be split or, under these constraints, overflow context. Requirement: check accumulated token usage and context limits. Dynamic chunking is not the direct fit.
Question 10
A document contains long tables. Fixed chunks separate row labels from values, making retrieved chunks ambiguous. Which approach should the developer use? Choose ONE.
- Top-k and metadata-filter review
- Chunking/preprocessing remediation
- Dynamic chunking
- Embedding-quality diagnostics
Correct Answer: C
Correct Answer
Answer C is correct because Dynamic chunking — for the described workload, adjusts document or context chunk, for the stated need, sizes so relevant content fits, for the stated need, model and retrieval constraints more, for the stated need, effectively. Best suited when fixed chunks are causing important, for this decision, passages to be split or, for this decision, overflow context. Deciding factor: choose chunking that preserves logically related table context. Dynamic chunking fits this case.
Incorrect Answers
Answer A is incorrect because Top-k and metadata-filter review — for the described workload, checks retrieval count and filters, for the stated need, to confirm that relevant documents, for the stated need, are not excluded or drowned, for the stated need, out. Best suited when the index is healthy but, for this decision, query configuration suppresses useful context. Deciding factor: choose chunking that preserves logically related table context. Top-k and metadata-filter review addresses something else.
Answer B is incorrect because Chunking/preprocessing remediation — for the described workload, fixes document splitting, normalization, metadata,, for the stated need, or preprocessing that prevents useful, for the stated need, passages from being retrieved. Best suited when source text is present but, for this decision, indexing preparation damages retrieval. Deciding factor: choose chunking that preserves logically related table context. Chunking/preprocessing remediation addresses something else.
Answer D is incorrect because Embedding-quality diagnostics — in the current architecture, examines whether semantically related content, in the described flow, is represented and retrieved as, in the described flow, expected by the embedding process. Applies where retrieval failure points to poor, in this case, vector representations rather than generation. Target outcome: choose chunking that preserves logically related table context. Embedding-quality diagnostics misses that outcome.
Question 11
The model returns a valid explanation but violates the requested ordering of required sections. What should the team implement? Choose ONE.
- Prompt design optimization
- Prompt test harness
- JSON Schema output validation
- Response-shape analysis
Correct Answer: A
Correct Answer
Answer A is correct because Prompt design optimization — for this operational need, restructures instructions, context ordering, delimiters,, under these constraints, or examples to make model, under these constraints, behavior more reliable. Architects use it when the model receives all necessary, for this workload, data but interprets the prompt, for this workload, poorly. Key need: strengthen prompt structure when content is present but instructions are not followed. Prompt design optimization aligns with it.
Incorrect Answers
Answer B is incorrect because Prompt test harness — for this operational need, runs controlled prompt cases repeatedly, under these constraints, so changes can be compared, under these constraints, against expected behaviors. Architects use it when prompt quality regressions need reproducible, for this workload, testing rather than ad hoc, for this workload, editing. Key need: strengthen prompt structure when content is present but instructions are not followed. Prompt test harness does not satisfy it.
Answer C is incorrect because JSON Schema output validation — in the current architecture, verifies that generated structured output, in the described flow, conforms to the required fields, in the described flow, and types before downstream processing. Applies where the model response must meet, in this case, a deterministic machine-readable contract. Target outcome: strengthen prompt structure when content is present but instructions are not followed. JSON Schema output validation misses that outcome.
Answer D is incorrect because Response-shape analysis — in the current architecture, examines returned status, metadata, stop, in the described flow, reason, and payload structure to, in the described flow, distinguish model behavior from integration, in the described flow, parsing errors. Applies where the caller receives a response, in this case, but handles it incorrectly. Target outcome: strengthen prompt structure when content is present but instructions are not followed. Response-shape analysis misses that outcome.
Question 12
Requests fail before reaching Bedrock because a serialized numeric field is sent as text and a required object is missing. Which option most directly addresses the requirement? Choose ONE.
- Request-schema validation
- Structured API error logging
- Parser/serialization check
- JSON Schema output validation
Correct Answer: A
Correct Answer
Answer A is correct because Request-schema validation — for this operational need, checks required fields, types, ranges,, under these constraints, and payload structure before sending, under these constraints, an FM request. Architects use it when invalid or malformed requests are, for this workload, reaching the model API. Key need: validate the outgoing request payload shape. Request-schema validation aligns with it.
Incorrect Answers
Answer B is incorrect because Structured API error logging — in this scenario, captures request identifiers, status codes,, in this case, exception details, and relevant non-sensitive, in this case, metadata for failed calls. Appropriate if integration failures must be diagnosed, at this layer, from concrete API evidence. The stem requires validate the outgoing request payload shape. Structured API error logging serves another purpose.
Answer C is incorrect because Parser/serialization check — under these constraints, verifies that the application correctly, at this layer, parses streamed or structured model, at this layer, responses without dropping fields or, at this layer, chunks. Choose it when the service returns data but, under these constraints, the application presents incomplete output. Requirement: validate the outgoing request payload shape. Parser/serialization check is not the direct fit.
Answer D is incorrect because JSON Schema output validation — for this operational need, verifies that generated structured output, under these constraints, conforms to the required fields, under these constraints, and types before downstream processing. Architects use it when the model response must meet, for this workload, a deterministic machine-readable contract. Key need: validate the outgoing request payload shape. JSON Schema output validation does not satisfy it.
Question 13
A response includes the expected data, but the application parser looks for a field name used by an older SDK version. Which design best satisfies this requirement? Choose ONE.
- Request-schema validation
- Response-shape analysis
- CloudWatch Logs correlation
- Parser/serialization check
Correct Answer: B
Correct Answer
Answer B is correct because Response-shape analysis — for this operational need, examines returned status, metadata, stop, under these constraints, reason, and payload structure to, under these constraints, distinguish model behavior from integration, under these constraints, parsing errors. Architects use it when the caller receives a response, for this workload, but handles it incorrectly. Key need: diagnose a caller-side mismatch with the returned response shape. Response-shape analysis aligns with it.
Incorrect Answers
Answer A is incorrect because Request-schema validation — within this implementation, checks required fields, types, ranges,, for this decision, and payload structure before sending, for this decision, an FM request. Strong fit when invalid or malformed requests are, within this design, reaching the model API. Operational need: diagnose a caller-side mismatch with the returned response shape. Request-schema validation is secondary here.
Answer C is incorrect because CloudWatch Logs correlation — within this implementation, centralizes application and service logs, for this decision, so request IDs and events, for this decision, can be correlated across failed, for this decision, interactions. Strong fit when troubleshooting requires a timeline of, within this design, logged events. Operational need: diagnose a caller-side mismatch with the returned response shape. CloudWatch Logs correlation is secondary here.
Answer D is incorrect because Parser/serialization check — for the described workload, verifies that the application correctly, for the stated need, parses streamed or structured model, for the stated need, responses without dropping fields or, for the stated need, chunks. Best suited when the service returns data but, for this decision, the application presents incomplete output. Deciding factor: diagnose a caller-side mismatch with the returned response shape. Parser/serialization check addresses something else.
Question 14
A new prompt template performs worse. The team wants exact side-by-side results from old and new versions on an unchanged evaluation set. What is the most appropriate implementation choice? Choose ONE.
- Systematic prompt refinement
- Prompt-version comparison
- Prompt test harness
- Model/configuration comparison
Correct Answer: B
Correct Answer
Answer B is correct because Prompt-version comparison — in the current architecture, compares outputs and metrics from, in the described flow, two prompt versions against the, in the described flow, same evaluation cases. Applies where a recent prompt change is, in this case, suspected of causing quality degradation. Target outcome: compare prompt revisions while holding test cases constant. Prompt-version comparison supplies it directly.
Incorrect Answers
Answer A is incorrect because Systematic prompt refinement is useful when the team intends to change prompt elements iteratively and measure whether quality improves. This requirement is different: compare the old and new templates exactly against the same evaluation set. Prompt-version comparison preserves the controlled baseline needed for a valid side-by-side result.
Answer C is incorrect because Prompt test harness — within this implementation, runs controlled prompt cases repeatedly, for this decision, so changes can be compared, for this decision, against expected behaviors. Strong fit when prompt quality regressions need reproducible, within this design, testing rather than ad hoc, within this design, editing. Operational need: compare prompt revisions while holding test cases constant. Prompt test harness is secondary here.
Answer D is incorrect because Model/configuration comparison — at this decision point, reproduces the same case with, within this design, a known-good model configuration to, within this design, isolate whether parameters or model, within this design, choice cause the symptom. Typical use: quality changed after model or, in the described flow, inference configuration changes. Goal: compare prompt revisions while holding test cases constant. Model/configuration comparison does not provide it.
Question 15
Vector search returns semantically weak matches after an embedding model migration, even though document text and filters are unchanged. Which approach should the developer use? Choose ONE.
- Embedding-quality diagnostics
- Retrieval relevance analysis
- Embedding drift monitoring
- Re-embed with consistent model
Correct Answer: A
Correct Answer
Answer A is correct because Embedding-quality diagnostics — for this operational need, examines whether semantically related content, under these constraints, is represented and retrieved as, under these constraints, expected by the embedding process. Architects use it when retrieval failure points to poor, for this workload, vector representations rather than generation. Key need: diagnose semantic representation quality after an embedding change. Embedding-quality diagnostics aligns with it.
Incorrect Answers
Answer B is incorrect because Retrieval relevance analysis — for the described workload, checks whether retrieved chunks actually, for the stated need, match the question and contain, for the stated need, the evidence the model needs. Best suited when RAG answers are wrong because, for this decision, supporting context may be irrelevant. Deciding factor: diagnose semantic representation quality after an embedding change. Retrieval relevance analysis addresses something else.
Answer C is incorrect because Embedding drift monitoring — in the current architecture, detects changes in retrieval behavior, in the described flow, or vector distributions after model,, in the described flow, corpus, or preprocessing changes. Applies where retrieval quality degrades gradually after, in this case, upstream changes. Target outcome: diagnose semantic representation quality after an embedding change. Embedding drift monitoring misses that outcome.
Answer D is incorrect because Re-embed with consistent model — at this decision point, recomputes vectors using the intended, within this design, embedding model and preprocessing so, within this design, query and corpus vectors are, within this design, compatible. Typical use: the index contains vectors built, in the described flow, with mismatched embedding settings. Goal: diagnose semantic representation quality after an embedding change. Re-embed with consistent model does not provide it.
Question 16
The query vectors use a new embedding model but the stored corpus vectors were never regenerated. What should the team implement? Choose ONE.
- Embedding drift monitoring
- Re-embed with consistent model
- Embedding-quality diagnostics
- Chunking/preprocessing remediation
Correct Answer: B
Correct Answer
Answer B is correct because Re-embed with consistent model — under these constraints, recomputes vectors using the intended, at this layer, embedding model and preprocessing so, at this layer, query and corpus vectors are, at this layer, compatible. Choose it when the index contains vectors built, under these constraints, with mismatched embedding settings. Requirement: make query and corpus vectors compatible. Re-embed with consistent model addresses it precisely.
Incorrect Answers
Answer A is incorrect because Embedding drift monitoring — for this operational need, detects changes in retrieval behavior, under these constraints, or vector distributions after model,, under these constraints, corpus, or preprocessing changes. Architects use it when retrieval quality degrades gradually after, for this workload, upstream changes. Key need: make query and corpus vectors compatible. Embedding drift monitoring does not satisfy it.
Answer C is incorrect because Embedding-quality diagnostics — within this implementation, examines whether semantically related content, for this decision, is represented and retrieved as, for this decision, expected by the embedding process. Strong fit when retrieval failure points to poor, within this design, vector representations rather than generation. Operational need: make query and corpus vectors compatible. Embedding-quality diagnostics is secondary here.
Answer D is incorrect because Chunking/preprocessing remediation — in the current architecture, fixes document splitting, normalization, metadata,, in the described flow, or preprocessing that prevents useful, in the described flow, passages from being retrieved. Applies where source text is present but, in this case, indexing preparation damages retrieval. Target outcome: make query and corpus vectors compatible. Chunking/preprocessing remediation misses that outcome.
Question 17
A relevant policy paragraph was split from its section title and metadata during preprocessing, causing poor retrieval. Which option most directly addresses the requirement? Choose ONE.
- Dynamic chunking
- Top-k and metadata-filter review
- Source-document freshness check
- Chunking/preprocessing remediation
Correct Answer: D
Correct Answer
Answer D is correct because Chunking/preprocessing remediation — for this operational need, fixes document splitting, normalization, metadata,, under these constraints, or preprocessing that prevents useful, under these constraints, passages from being retrieved. Architects use it when source text is present but, for this workload, indexing preparation damages retrieval. Key need: repair preprocessing that removed useful structure and metadata. Chunking/preprocessing remediation aligns with it.
Incorrect Answers
Answer A is incorrect because Dynamic chunking — in the current architecture, adjusts document or context chunk, in the described flow, sizes so relevant content fits, in the described flow, model and retrieval constraints more, in the described flow, effectively. Applies where fixed chunks are causing important, in this case, passages to be split or, in this case, overflow context. Target outcome: repair preprocessing that removed useful structure and metadata. Dynamic chunking misses that outcome.
Answer B is incorrect because Top-k and metadata-filter review — in the current architecture, checks retrieval count and filters, in the described flow, to confirm that relevant documents, in the described flow, are not excluded or drowned, in the described flow, out. Applies where the index is healthy but, in this case, query configuration suppresses useful context. Target outcome: repair preprocessing that removed useful structure and metadata. Top-k and metadata-filter review misses that outcome.
Answer C is incorrect because Source-document freshness check — in the current architecture, confirms the retrieval corpus contains, in the described flow, the current source content and, in the described flow, that updates were indexed. Applies where answers use outdated facts because, in this case, the knowledge source may be, in this case, stale. Target outcome: repair preprocessing that removed useful structure and metadata. Source-document freshness check misses that outcome.
Question 18
A vector query is accurate but takes several seconds because search parameters favor unnecessary exhaustive work. Which design best satisfies this requirement? Choose ONE.
- AWS X-Ray tracing
- Retrieval relevance analysis
- Vector-search performance tuning
- Top-k and metadata-filter review
Correct Answer: C
Correct Answer
Answer C is correct because Vector-search performance tuning — under these constraints, optimizes vector index/search parameters and, at this layer, query patterns to reduce latency, at this layer, while preserving acceptable recall. Choose it when retrieval is relevant but too slow. Requirement: reduce vector search latency while preserving acceptable retrieval quality. Vector-search performance tuning addresses it precisely.
Incorrect Answers
Answer A is incorrect because AWS X-Ray can trace distributed request paths and reveal where latency occurs across services. In this case, the slow component is already known: the vector query is doing unnecessary exhaustive search work. Vector-search performance tuning changes the search parameters causing the delay; tracing would diagnose rather than fix that known configuration problem.
Answer B is incorrect because Retrieval relevance analysis — in the current architecture, checks whether retrieved chunks actually, in the described flow, match the question and contain, in the described flow, the evidence the model needs. Applies where RAG answers are wrong because, in this case, supporting context may be irrelevant. Target outcome: reduce vector search latency while preserving acceptable retrieval quality. Retrieval relevance analysis misses that outcome.
Answer D is incorrect because Top-k and metadata-filter review — for this operational need, checks retrieval count and filters, under these constraints, to confirm that relevant documents, under these constraints, are not excluded or drowned, under these constraints, out. Architects use it when the index is healthy but, for this workload, query configuration suppresses useful context. Key need: reduce vector search latency while preserving acceptable retrieval quality. Top-k and metadata-filter review does not satisfy it.
Question 19
An operations team enabled Amazon Bedrock model invocation logging, but OpenAI-compatible Responses API calls sent through the bedrock-mantle endpoint never appear in the invocation logs. The team must keep the OpenAI-compatible API style and make these requests eligible for Bedrock invocation logging. What should they change? Choose ONE.
- Publish the invocation logs only to Amazon S3
- Enable CloudTrail management events for Amazon Bedrock
- Increase the AWS X-Ray sampling rate
- Send the calls through the bedrock-runtime endpoint
Correct Answer: D
Correct Answer
Answer D is correct because Amazon Bedrock model invocation logging currently supports calls made through the bedrock-runtime endpoint, including the OpenAI-compatible Responses and Chat Completions APIs on that endpoint. Moving these requests to bedrock-runtime makes them eligible for invocation logging while preserving the API style. The issue is endpoint support, not trace sampling or the log destination.
Incorrect Answers
Answer A is incorrect because Amazon S3 is a supported destination for Bedrock invocation logs, but changing destinations does not make an unsupported invocation endpoint eligible for logging. If the call is made through bedrock-mantle, the logging limitation remains. The team should first route the OpenAI-compatible request through bedrock-runtime, then retain its configured CloudWatch Logs or Amazon S3 destination.
Answer B is incorrect because AWS CloudTrail can provide an audit history of AWS API activity, which is useful for governance and investigation. It is not a substitute for Amazon Bedrock model invocation logging that captures supported invocation request, response, and metadata records. Enabling CloudTrail management events therefore does not resolve the endpoint-specific absence of Bedrock invocation logs.
Answer C is incorrect because AWS X-Ray sampling controls how many distributed traces are recorded. It does not change which Amazon Bedrock endpoints are supported by model invocation logging. Raising the sample rate could produce more traces, but calls made through an unsupported endpoint still would not become Bedrock invocation-log records. The required fix is to use the supported runtime endpoint.
Question 20
Generated JSON sometimes omits a required field even though the text is otherwise sensible; downstream code must reject invalid payloads. Which approach should the developer use? Choose ONE.
- Response-shape analysis
- JSON Schema output validation
- Prompt design optimization
- Request-schema validation
Correct Answer: B
Correct Answer
Answer B is correct because JSON Schema output validation — within this implementation, verifies that generated structured output, for this decision, conforms to the required fields, for this decision, and types before downstream processing. Strong fit when the model response must meet, within this design, a deterministic machine-readable contract. Operational need: enforce required fields and types on model output. JSON Schema output validation is appropriate here.
Incorrect Answers
Answer A is incorrect because Response-shape analysis — within this implementation, examines returned status, metadata, stop, for this decision, reason, and payload structure to, for this decision, distinguish model behavior from integration, for this decision, parsing errors. Strong fit when the caller receives a response, within this design, but handles it incorrectly. Operational need: enforce required fields and types on model output. Response-shape analysis is secondary here.
Answer C is incorrect because Prompt design optimization — within this implementation, restructures instructions, context ordering, delimiters,, for this decision, or examples to make model, for this decision, behavior more reliable. Strong fit when the model receives all necessary, within this design, data but interprets the prompt, within this design, poorly. Operational need: enforce required fields and types on model output. Prompt design optimization is secondary here.
Answer D is incorrect because Request-schema validation — in this scenario, checks required fields, types, ranges,, in this case, and payload structure before sending, in this case, an FM request. Appropriate if invalid or malformed requests are, at this layer, reaching the model API. The stem requires enforce required fields and types on model output. Request-schema validation serves another purpose.
