免费 AWS 认证生成式 AI 开发者考试练习

Answer: A

Why:

This approach provides the best balance of cost, latency, and flexibility. Dynamic model selection using AWS AppConfig allows the application to route most requests to a lower-cost, latency-optimized model while escalating complex cases to a higher-capability model without code changes. Cross-Region Inference improves availability during regional disruptions, and on-demand Bedrock invocation avoids paying for idle capacity during non-peak periods.

Why not:

Option B: SageMaker real-time endpoints introduce fixed GPU costs and operational overhead that are inefficient for bursty, text-based workloads already supported by managed Bedrock models.
Option C: Provisioned throughput guarantees latency but significantly increases cost by reserving peak capacity at all times, even during off-peak periods.
Option D: Knowledge Bases are designed for retrieval augmentation and grounding, not for real-time analysis of short, incoming alerts, and would add unnecessary retrieval latency.

Answer: A

Why:

This approach provides the best balance of cost, latency, and flexibility. Dynamic model selection using AWS AppConfig allows the application to route most requests to a lower-cost, latency-optimized model while escalating complex cases to a higher-capability model without code changes. Cross-Region Inference improves availability during regional disruptions, and on-demand Bedrock invocation avoids paying for idle capacity during non-peak periods.

Why not:

Option B: SageMaker real-time endpoints introduce fixed GPU costs and operational overhead that are inefficient for bursty, text-based workloads already supported by managed Bedrock models.
Option C: Provisioned throughput guarantees latency but significantly increases cost by reserving peak capacity at all times, even during off-peak periods.
Option D: Knowledge Bases are designed for retrieval augmentation and grounding, not for real-time analysis of short, incoming alerts, and would add unnecessary retrieval latency.

Answers: A, C, and E

Why:

Option A: Segmenting content into multiple indexes by document domain helps reduce search scope and improve retrieval performance at a large scale. With more than 12 million documents spanning very different content types, domain-based index organization can improve relevance and support faster queries by avoiding unnecessary cross-domain retrieval. This design choice complements the use of OpenSearch Serverless.
Option C: Amazon Bedrock Knowledge Bases backed by Amazon OpenSearch Serverless provide a managed approach for semantic and hybrid retrieval with minimal operational overhead. This aligns well with the requirement to support semantic search at scale while reducing infrastructure management burden. Hybrid search can improve retrieval accuracy by combining vector similarity with keyword matching.
Option E: A hierarchical metadata framework using attributes such as region, document type, and update timestamp enables efficient filtering and improves retrieval precision. This is especially important in a large enterprise knowledge system where users need only relevant subsets of documents returned within strict latency targets.

Why not:

Option B: Aurora PostgreSQL with pgvector can work for smaller or moderate-scale vector workloads, but for a corpus of over 12 million documents with strict latency targets, it typically requires more tuning, scaling effort, and operational management than a purpose-built managed search solution.
Option D: DynamoDB is not designed for large-scale vector similarity search, and computing cosine similarity in Lambda for each query would be inefficient and unlikely to meet the required latency at this scale.
Option F: Using only fixed-size chunking can reduce retrieval quality because it often splits content at arbitrary boundaries rather than preserving document meaning. This can separate related paragraphs, break tables or section context, and mix unrelated ideas into a single chunk, which makes retrieved passages less precise and less useful for downstream generation.

Answers: A, C, and E

Why:

Option A: Segmenting content into multiple indexes by document domain helps reduce search scope and improve retrieval performance at a large scale. With more than 12 million documents spanning very different content types, domain-based index organization can improve relevance and support faster queries by avoiding unnecessary cross-domain retrieval. This design choice complements the use of OpenSearch Serverless.
Option C: Amazon Bedrock Knowledge Bases backed by Amazon OpenSearch Serverless provide a managed approach for semantic and hybrid retrieval with minimal operational overhead. This aligns well with the requirement to support semantic search at scale while reducing infrastructure management burden. Hybrid search can improve retrieval accuracy by combining vector similarity with keyword matching.
Option E: A hierarchical metadata framework using attributes such as region, document type, and update timestamp enables efficient filtering and improves retrieval precision. This is especially important in a large enterprise knowledge system where users need only relevant subsets of documents returned within strict latency targets.

Why not:

Option B: Aurora PostgreSQL with pgvector can work for smaller or moderate-scale vector workloads, but for a corpus of over 12 million documents with strict latency targets, it typically requires more tuning, scaling effort, and operational management than a purpose-built managed search solution.
Option D: DynamoDB is not designed for large-scale vector similarity search, and computing cosine similarity in Lambda for each query would be inefficient and unlikely to meet the required latency at this scale.
Option F: Using only fixed-size chunking can reduce retrieval quality because it often splits content at arbitrary boundaries rather than preserving document meaning. This can separate related paragraphs, break tables or section context, and mix unrelated ideas into a single chunk, which makes retrieved passages less precise and less useful for downstream generation.

Permutation Visualization

Why:

Audio must first be converted to text before any validation or enrichment can occur. Data quality validation ensures only usable transcripts proceed further in the pipeline. Generating embeddings from raw or unvalidated transcripts would reduce search accuracy and degrade RAG performance. Entity extraction improves downstream retrieval relevance by enriching metadata. Finally, embeddings are generated only after the text is validated and enriched, ensuring high-quality vectors for RAG.

Why:

Audio must first be converted to text before any validation or enrichment can occur. Data quality validation ensures only usable transcripts proceed further in the pipeline. Generating embeddings from raw or unvalidated transcripts would reduce search accuracy and degrade RAG performance. Entity extraction improves downstream retrieval relevance by enriching metadata. Finally, embeddings are generated only after the text is validated and enriched, ensuring high-quality vectors for RAG.

Answer: B

Why:
Bedrock Prompt Management provides native prompt versioning, approval workflows, parameterized templates, and audit logging, which directly addresses the requirements for controlled prompt updates, traceability, and enterprise governance. This approach ensures that the central AI governance team can safely update prompts without breaking downstream applications while maintaining full regulatory compliance.

Why not:

Option A: S3 versioning lacks approval workflows and structured governance controls.
Option C: Embedding prompts in code tightly couples prompt changes to deployments, slowing iteration and increasing risk.
Option D: DynamoDB requires custom governance logic and auditing mechanisms.

Answer: B

Why:
Bedrock Prompt Management provides native prompt versioning, approval workflows, parameterized templates, and audit logging, which directly addresses the requirements for controlled prompt updates, traceability, and enterprise governance. This approach ensures that the central AI governance team can safely update prompts without breaking downstream applications while maintaining full regulatory compliance.

Why not:

Option A: S3 versioning lacks approval workflows and structured governance controls.
Option C: Embedding prompts in code tightly couples prompt changes to deployments, slowing iteration and increasing risk.
Option D: DynamoDB requires custom governance logic and auditing mechanisms.

Answer: A

Why:
Amazon Bedrock Prompt Flows natively support multi-step prompt chaining, conditional execution, and reusable components, making them ideal for complex analytical reasoning workflows. Integrating Prompt Management enables versioning, approvals, and governance, while Step Functions support systematic testing and regression validation. This approach aligns with enterprise prompt governance best practices and minimizes custom orchestration logic.

Why not:

Option B: S3 and Lambda require extensive custom logic and lack native governance and testing features.
Option C: A single monolithic prompt is brittle, difficult to govern, and hard to evolve safely.
Option D: DynamoDB-based prompt assembly increases complexity and does not provide built-in approval or flow control.

Answer: A

Why:
Amazon Bedrock Prompt Flows natively support multi-step prompt chaining, conditional execution, and reusable components, making them ideal for complex analytical reasoning workflows. Integrating Prompt Management enables versioning, approvals, and governance, while Step Functions support systematic testing and regression validation. This approach aligns with enterprise prompt governance best practices and minimizes custom orchestration logic.

Why not:

Option B: S3 and Lambda require extensive custom logic and lack native governance and testing features.
Option C: A single monolithic prompt is brittle, difficult to govern, and hard to evolve safely.
Option D: DynamoDB-based prompt assembly increases complexity and does not provide built-in approval or flow control.

Answer: C

Why:
Invoking Amazon Bedrock models directly from AWS Lambda with streaming APIs provides fully managed, elastic scaling without capacity planning, making it ideal for bursty workloads. Lambda automatically scales with concurrent requests, and Bedrock streaming APIs allow real-time token delivery to meet the sub-300 ms first-token latency requirement. This approach minimizes cost during idle periods because there is no provisioned capacity to pay for. It also avoids operational overhead associated with managing GPU instances or containers.

Why not:

Option A: Provisioned throughput ensures consistent capacity but is cost-inefficient when traffic is unpredictable and idle for long durations.
Option B: SageMaker real-time endpoints incur continuous costs for provisioned GPU instances and require careful scaling configuration, which is inefficient for highly spiky workloads with long idle periods.
Option D: ECS with Fargate introduces additional operational complexity and cold-start latency, and managing streaming over ALB requires more custom infrastructure than necessary.

Answer: C

Why:
Invoking Amazon Bedrock models directly from AWS Lambda with streaming APIs provides fully managed, elastic scaling without capacity planning, making it ideal for bursty workloads. Lambda automatically scales with concurrent requests, and Bedrock streaming APIs allow real-time token delivery to meet the sub-300 ms first-token latency requirement. This approach minimizes cost during idle periods because there is no provisioned capacity to pay for. It also avoids operational overhead associated with managing GPU instances or containers.

Why not:

Option A: Provisioned throughput ensures consistent capacity but is cost-inefficient when traffic is unpredictable and idle for long durations.
Option B: SageMaker real-time endpoints incur continuous costs for provisioned GPU instances and require careful scaling configuration, which is inefficient for highly spiky workloads with long idle periods.
Option D: ECS with Fargate introduces additional operational complexity and cold-start latency, and managing streaming over ALB requires more custom infrastructure than necessary.

Answers: B and D

Why:

Option B: Amazon Bedrock Agents natively support ReAct-style reasoning, tool invocation with parameter validation, and minimize custom orchestration code.

Option D: AWS Step Functions provides reliable orchestration for multi-step reasoning, tool execution, and human-in-the-loop escalation with built-in retries and state handling.

Why not:

Option A: Amazon S3 is unsuitable for low-latency, frequently updated conversational state.
Option C: Lambda-based MCP servers help expose tools but do not handle orchestration or state management on their own.

Option E: ECS/Fargate significantly increases operational overhead and duplicates capabilities already provided by Bedrock Agents and Step Functions.

Answers: B and D

Why:

Option B: Amazon Bedrock Agents natively support ReAct-style reasoning, tool invocation with parameter validation, and minimize custom orchestration code.

Option D: AWS Step Functions provides reliable orchestration for multi-step reasoning, tool execution, and human-in-the-loop escalation with built-in retries and state handling.

Why not:

Option A: Amazon S3 is unsuitable for low-latency, frequently updated conversational state.
Option C: Lambda-based MCP servers help expose tools but do not handle orchestration or state management on their own.

Option E: ECS/Fargate significantly increases operational overhead and duplicates capabilities already provided by Bedrock Agents and Step Functions.

Permutation Visualization

Why:
Version control is the foundation for traceability and auditability. Automated testing and security scans must occur before any deployment to catch regressions or vulnerabilities early. Deploying to a staging environment allows validation in isolation before impacting users. Finally, controlled promotion with rollback ensures resilience and safe production releases.

Why:
Version control is the foundation for traceability and auditability. Automated testing and security scans must occur before any deployment to catch regressions or vulnerabilities early. Deploying to a staging environment allows validation in isolation before impacting users. Finally, controlled promotion with rollback ensures resilience and safe production releases.

Match the Answers

Why:

MCP servers, implemented via AWS Lambda and Amazon ECS, provide a critical abstraction layer for tool integration. They enforce least-privilege access with IAM roles and supply a standardized schema for LLM interactions. Lambda handles short, event-driven tasks, while ECS manages long-running, compute-intensive processes. AWS X-Ray enables end-to-end request tracing across agent workflows, visualizing bottlenecks and failures. AWS Step Functions orchestrate retries and fallback behavior, ensuring workflows gracefully handle downstream failures. Direct API exposure is an anti-pattern here, as it lacks the validation and security needed to protect internal microservices from unvetted LLM requests.

Why:

MCP servers, implemented via AWS Lambda and Amazon ECS, provide a critical abstraction layer for tool integration. They enforce least-privilege access with IAM roles and supply a standardized schema for LLM interactions. Lambda handles short, event-driven tasks, while ECS manages long-running, compute-intensive processes. AWS X-Ray enables end-to-end request tracing across agent workflows, visualizing bottlenecks and failures. AWS Step Functions orchestrate retries and fallback behavior, ensuring workflows gracefully handle downstream failures. Direct API exposure is an anti-pattern here, as it lacks the validation and security needed to protect internal microservices from unvetted LLM requests.

Answer: A

Why:
Step Functions provide explicit control over parallel execution, timeouts, retries, and fallback logic, which is essential for graceful degradation. By isolating agents and managing the state externally in DynamoDB, the system can skip failed components and still return partial results safely. This approach follows best practices for resilience and fault isolation in distributed GenAI systems. It ensures predictable behavior under failure conditions and meets strict latency requirements.

Why not:

Option B: Prompt-based failure handling is non-deterministic and cannot guarantee timely or safe degradation.
Option C: ECS-based synchronous coordination increases latency and operational complexity, making it harder to meet the 2-second response target.
Option D: A monolithic agent lacks fault isolation, becoming a single point of failure.

Answer: A

Why:
Step Functions provide explicit control over parallel execution, timeouts, retries, and fallback logic, which is essential for graceful degradation. By isolating agents and managing the state externally in DynamoDB, the system can skip failed components and still return partial results safely. This approach follows best practices for resilience and fault isolation in distributed GenAI systems. It ensures predictable behavior under failure conditions and meets strict latency requirements.

Why not:

Option B: Prompt-based failure handling is non-deterministic and cannot guarantee timely or safe degradation.
Option C: ECS-based synchronous coordination increases latency and operational complexity, making it harder to meet the 2-second response target.
Option D: A monolithic agent lacks fault isolation, becoming a single point of failure.

Answer: C

Why:
Amazon Bedrock Guardrails provide native, model-level enforcement of content safety policies, including PII filtering and denied topics, which is essential for HIPAA-regulated workloads. Applying guardrails at inference time ensures consistent enforcement across all requests without relying on model compliance alone. This approach balances strong protection with usability and minimizes operational overhead by avoiding custom post-processing logic. It aligns with defense-in-depth and responsible AI principles by preventing violations before responses are delivered.

Why not:

Option A: Prompt-based instructions are non-deterministic and vulnerable to prompt injection or jailbreak attempts.

Option B: Post-generation redaction reduces risk but still allows sensitive content to be generated internally, which increases exposure and does not fully satisfy HIPAA’s preventive controls.

Option D: Offline auditing detects violations after the fact and does not prevent real-time disclosure, which is unacceptable for low-risk-tolerance health care systems.

Answer: C

Why:
Amazon Bedrock Guardrails provide native, model-level enforcement of content safety policies, including PII filtering and denied topics, which is essential for HIPAA-regulated workloads. Applying guardrails at inference time ensures consistent enforcement across all requests without relying on model compliance alone. This approach balances strong protection with usability and minimizes operational overhead by avoiding custom post-processing logic. It aligns with defense-in-depth and responsible AI principles by preventing violations before responses are delivered.

Why not:

Option A: Prompt-based instructions are non-deterministic and vulnerable to prompt injection or jailbreak attempts.

Option B: Post-generation redaction reduces risk but still allows sensitive content to be generated internally, which increases exposure and does not fully satisfy HIPAA’s preventive controls.

Option D: Offline auditing detects violations after the fact and does not prevent real-time disclosure, which is unacceptable for low-risk-tolerance health care systems.

Answers: B and E

Why:

Option B: Bedrock Guardrails enforce real-time PII redaction at inference, directly supporting GDPR data minimization and output safety.

Option E: KMS-managed encryption and restrictive S3 bucket policies ensure strong encryption at rest and least-privilege access, both core GDPR requirements.

Why not:

Option A: Amazon Macie detects PII exposure but does not prevent PII leakage during inference.

Option C: Network isolation improves security posture but does not address PII protection or encryption by itself.

Option D: Disabling encryption violates GDPR security obligations.

Answers: B and E

Why:

Option B: Bedrock Guardrails enforce real-time PII redaction at inference, directly supporting GDPR data minimization and output safety.

Option E: KMS-managed encryption and restrictive S3 bucket policies ensure strong encryption at rest and least-privilege access, both core GDPR requirements.

Why not:

Option A: Amazon Macie detects PII exposure but does not prevent PII leakage during inference.

Option C: Network isolation improves security posture but does not address PII protection or encryption by itself.

Option D: Disabling encryption violates GDPR security obligations.

Permutation Visualization

Why:
Input sanitization must occur first to reduce adversarial risk before classification. Content classification then identifies sensitive or toxic inputs early. Bedrock Guardrails enforce model-level safety during generation, and post-processing validation provides a final checkpoint before responses reach users. This layered approach aligns with defense-in-depth principles required in government systems.

Why:
Input sanitization must occur first to reduce adversarial risk before classification. Content classification then identifies sensitive or toxic inputs early. Bedrock Guardrails enforce model-level safety during generation, and post-processing validation provides a final checkpoint before responses reach users. This layered approach aligns with defense-in-depth principles required in government systems.

Answer: A

Why:
SageMaker model cards are designed to improve transparency and responsible AI governance by documenting the model’s intended use, known limitations, evaluation results, and performance considerations, and governance context. This directly supports the requirement to make model behavior understandable and appropriately governed in a low-risk, highly regulated environment. Model cards are specifically aligned to responsible AI principles because they help stakeholders understand when and how a model should be used, along with its constraints.

Why not:

Option B: CloudTrail and model invocation logging are important for auditability and compliance because they help track who used the model, when it was invoked, and what actions occurred. However, they do not primarily document model limitations, intended use, or evaluation characteristics. This makes B a strong governance control, but not the best answer for transparency about model behavior.

Option C: Response citations improve traceability and explainability of retrieved data sources in a RAG workflow. They help analysts verify which documents informed an answer, which is valuable for trust and provenance. However, citations do not explain the model’s intended use, weaknesses, or evaluation results, so they are not the best fit for the question’s focus on responsible AI transparency.

Option D: A formal approval workflow strengthens change management and operational governance by controlling modifications to prompts and retrieval settings. This is useful in regulated environments, but it does not by itself provide transparency into the model’s capabilities, limitations, or evaluation outcomes. It is therefore a governance control, not the strongest transparency mechanism.

Answer: A

Why:
SageMaker model cards are designed to improve transparency and responsible AI governance by documenting the model’s intended use, known limitations, evaluation results, and performance considerations, and governance context. This directly supports the requirement to make model behavior understandable and appropriately governed in a low-risk, highly regulated environment. Model cards are specifically aligned to responsible AI principles because they help stakeholders understand when and how a model should be used, along with its constraints.

Why not:

Option B: CloudTrail and model invocation logging are important for auditability and compliance because they help track who used the model, when it was invoked, and what actions occurred. However, they do not primarily document model limitations, intended use, or evaluation characteristics. This makes B a strong governance control, but not the best answer for transparency about model behavior.

Option C: Response citations improve traceability and explainability of retrieved data sources in a RAG workflow. They help analysts verify which documents informed an answer, which is valuable for trust and provenance. However, citations do not explain the model’s intended use, weaknesses, or evaluation results, so they are not the best fit for the question’s focus on responsible AI transparency.

Option D: A formal approval workflow strengthens change management and operational governance by controlling modifications to prompts and retrieval settings. This is useful in regulated environments, but it does not by itself provide transparency into the model’s capabilities, limitations, or evaluation outcomes. It is therefore a governance control, not the strongest transparency mechanism.

Answer: C

Why:
Combining input sanitization, intent validation, and Bedrock Guardrails provides layered protection against prompt injection without eliminating legitimate user queries. Guardrails prevent disclosure of system instructions, while CloudWatch metrics enable continuous monitoring and improvement. This approach aligns with responsible AI by preserving utility while reducing adversarial risk.

Why not:

Option A: Temperature tuning does not reliably prevent prompt injection and can increase unpredictability.

Option B: Over-restricting input severely degrades usability and business value.

Option D: Post-generation filtering allows sensitive information to be generated internally and does not prevent exploitation.

Answer: C

Why:
Combining input sanitization, intent validation, and Bedrock Guardrails provides layered protection against prompt injection without eliminating legitimate user queries. Guardrails prevent disclosure of system instructions, while CloudWatch metrics enable continuous monitoring and improvement. This approach aligns with responsible AI by preserving utility while reducing adversarial risk.

Why not:

Option A: Temperature tuning does not reliably prevent prompt injection and can increase unpredictability.

Option B: Over-restricting input severely degrades usability and business value.

Option D: Post-generation filtering allows sensitive information to be generated internally and does not prevent exploitation.

Answer: A

Why:
Amazon Bedrock Prompt Caching is well-suited for workloads in which many requests share large, identical prompt prefixes. By caching those repeated input tokens, Bedrock reduces the amount of prompt content that must be reprocessed on each request, lowering inference cost and often improving latency. This approach preserves model quality because it does not require changing models or simplifying responses, making it the best fit for reducing cost without materially degrading output quality.

Why not:
Option B: Using the smallest model for every request may lower cost, but it risks unacceptable quality loss for more complex questions.
Option C: SageMaker AI real-time endpoints increase operational overhead and fixed infrastructure cost, and do not directly optimize the Bedrock token-processing cost pattern described.
Option D: Temperature affects randomness and style, not prompt token processing, so it is not a dependable cost-reduction mechanism.

Answer: A

Why:
Amazon Bedrock Prompt Caching is well-suited for workloads in which many requests share large, identical prompt prefixes. By caching those repeated input tokens, Bedrock reduces the amount of prompt content that must be reprocessed on each request, lowering inference cost and often improving latency. This approach preserves model quality because it does not require changing models or simplifying responses, making it the best fit for reducing cost without materially degrading output quality.

Why not:
Option B: Using the smallest model for every request may lower cost, but it risks unacceptable quality loss for more complex questions.
Option C: SageMaker AI real-time endpoints increase operational overhead and fixed infrastructure cost, and do not directly optimize the Bedrock token-processing cost pattern described.
Option D: Temperature affects randomness and style, not prompt token processing, so it is not a dependable cost-reduction mechanism.

Answers: A, B, and E

Why:

Option A: Semantic caching reduces repeated inference while maintaining relevance, improving latency, and cost efficiency.

Option B: Streaming responses significantly reduce perceived latency by delivering tokens incrementally, meeting the first-token SLA.

Option E: Pre-computation shifts expensive inference to off-peak hours and enables fast lookups during peak traffic.

Why not:

Option C: Batching increases throughput but can increase per-request latency, violating the first-token SLA.

Option D: Higher-dimensional embeddings increase compute cost and query latency without guaranteed performance gains.

Option F: Edge caching ignores personalization and risks serving incorrect responses to users.

Answers: A, B, and E

Why:

Option A: Semantic caching reduces repeated inference while maintaining relevance, improving latency, and cost efficiency.

Option B: Streaming responses significantly reduce perceived latency by delivering tokens incrementally, meeting the first-token SLA.

Option E: Pre-computation shifts expensive inference to off-peak hours and enables fast lookups during peak traffic.

Why not:

Option C: Batching increases throughput but can increase per-request latency, violating the first-token SLA.

Option D: Higher-dimensional embeddings increase compute cost and query latency without guaranteed performance gains.

Option F: Edge caching ignores personalization and risks serving incorrect responses to users.

Permutation Visualization

Why:
Model invocation logs provide the raw data needed for analysis. Custom metrics translate logs into actionable signals. Dashboards and alarms enable proactive monitoring and alerting. Distributed tracing with X-Ray then adds deep diagnostic capability for correlating performance issues across services.

Why:
Model invocation logs provide the raw data needed for analysis. Custom metrics translate logs into actionable signals. Dashboards and alarms enable proactive monitoring and alerting. Distributed tracing with X-Ray then adds deep diagnostic capability for correlating performance issues across services.

Match the Answers

Why:
Model cascading aligns model cost with query complexity. Streaming reduces first-token latency without impacting accuracy. LLM-as-a-judge enables scalable hallucination monitoring. Anomaly detection identifies unexpected operational behavior in production.

Why:
Model cascading aligns model cost with query complexity. Streaming reduces first-token latency without impacting accuracy. LLM-as-a-judge enables scalable hallucination monitoring. Anomaly detection identifies unexpected operational behavior in production.

Answer: A

Why:
CloudWatch anomaly detection provides near real-time identification of unusual latency or token-usage patterns, enabling automated response during unpredictable traffic spikes. By switching only among pre-validated model tiers that already meet the required factual accuracy threshold, the company can control cost and latency without violating quality requirements. This approach best satisfies the need for automated anomaly handling while maintaining performance and accuracy objectives.

Why not:

Option B: AWS Cost Explorer and AWS Budgets can help identify spending trends and projected overruns, but this approach still depends on manual review and manual intervention. That makes it too slow for unpredictable breaking-news spikes, where the system needs automated detection and immediate response to protect both latency and cost targets.

Option C: Provisioning capacity year-round for worst-case traffic may help maintain latency and preserve quality by keeping the highest-accuracy model available, but it is expensive and inefficient during normal demand. It is also a static approach that does not provide anomaly detection or adaptive optimization to balance cost, latency, and accuracy dynamically.

Option D: Disabling streaming and standardizing response handling may simplify observability and reduce some response variability, but it does not address the main problem of unpredictable traffic spikes or provide meaningful cost control. It also does not improve factual accuracy, and disabling streaming can worsen perceived responsiveness for end users.

Answer: A

Why:
CloudWatch anomaly detection provides near real-time identification of unusual latency or token-usage patterns, enabling automated response during unpredictable traffic spikes. By switching only among pre-validated model tiers that already meet the required factual accuracy threshold, the company can control cost and latency without violating quality requirements. This approach best satisfies the need for automated anomaly handling while maintaining performance and accuracy objectives.

Why not:

Option B: AWS Cost Explorer and AWS Budgets can help identify spending trends and projected overruns, but this approach still depends on manual review and manual intervention. That makes it too slow for unpredictable breaking-news spikes, where the system needs automated detection and immediate response to protect both latency and cost targets.

Option C: Provisioning capacity year-round for worst-case traffic may help maintain latency and preserve quality by keeping the highest-accuracy model available, but it is expensive and inefficient during normal demand. It is also a static approach that does not provide anomaly detection or adaptive optimization to balance cost, latency, and accuracy dynamically.

Option D: Disabling streaming and standardizing response handling may simplify observability and reduce some response variability, but it does not address the main problem of unpredictable traffic spikes or provide meaningful cost control. It also does not improve factual accuracy, and disabling streaming can worsen perceived responsiveness for end users.

Answer: D

Why:
Amazon Bedrock Model Evaluations with LLM-as-a-judge provide an automated and repeatable way to assess factual accuracy and relevance, which directly addresses the observed omissions. Using a golden dataset ensures consistent comparisons across prompt versions. This approach scales without heavy human involvement while still producing interpretable scores aligned with GenAI quality dimensions. It is specifically designed for GenAI evaluation rather than surface-level text similarity.

Why not:

Option A: ROUGE and BLEU measure textual overlap, not factual accuracy or hallucination behavior, making them insufficient for GenAI quality assurance.

Option B: Latency and length metrics do not measure factual correctness or omission of key information.

Option C: Manual feedback does not scale and lacks consistency for regression testing.

Answer: D

Why:
Amazon Bedrock Model Evaluations with LLM-as-a-judge provide an automated and repeatable way to assess factual accuracy and relevance, which directly addresses the observed omissions. Using a golden dataset ensures consistent comparisons across prompt versions. This approach scales without heavy human involvement while still producing interpretable scores aligned with GenAI quality dimensions. It is specifically designed for GenAI evaluation rather than surface-level text similarity.

Why not:

Option A: ROUGE and BLEU measure textual overlap, not factual accuracy or hallucination behavior, making them insufficient for GenAI quality assurance.

Option B: Latency and length metrics do not measure factual correctness or omission of key information.

Option C: Manual feedback does not scale and lacks consistency for regression testing.

Answers: B and C

Why:

Option B is correct because retrieval relevance scoring directly validates whether the RAG retriever is returning the correct documents, a primary cause of hallucinations.

Option C is correct because LLM-as-a-judge allows separation of retrieval quality issues from generation issues, enabling precise root-cause isolation.

Why not:

Option A: Embedding drift analysis is indirect and insufficient without relevance validation.

Option D: Token and length metrics do not explain hallucinations or factual accuracy drops.

Option E: Temperature changes introduce noise and do not diagnose retrieval failures.

Answers: B and C

Why:

Option B is correct because retrieval relevance scoring directly validates whether the RAG retriever is returning the correct documents, a primary cause of hallucinations.

Option C is correct because LLM-as-a-judge allows separation of retrieval quality issues from generation issues, enabling precise root-cause isolation.

Why not:

Option A: Embedding drift analysis is indirect and insufficient without relevance validation.

Option D: Token and length metrics do not explain hallucinations or factual accuracy drops.

Option E: Temperature changes introduce noise and do not diagnose retrieval failures.

Permutation Visualization

Why:
Reviewing recent changes first narrows the scope of the investigation. Verifying embedding consistency detects common causes of silent retrieval failure, such as embedding drift. Comparing against a golden dataset confirms whether retrieval relevance has degraded. Finally, log analysis provides detailed evidence to validate findings and identify secondary issues.

Why:
Reviewing recent changes first narrows the scope of the investigation. Verifying embedding consistency detects common causes of silent retrieval failure, such as embedding drift. Comparing against a golden dataset confirms whether retrieval relevance has degraded. Finally, log analysis provides detailed evidence to validate findings and identify secondary issues.

Match the Answers

Why:
Amazon Bedrock Agent evaluations provide agent-level task completion and tool usage metrics. Canary deployments orchestrated with Step Functions prevent regressions from reaching production. DynamoDB is the appropriate low-latency store for agent memory across steps. CloudWatch Logs are useful diagnostically, but do not directly satisfy the listed requirements.

Why:
Amazon Bedrock Agent evaluations provide agent-level task completion and tool usage metrics. Canary deployments orchestrated with Step Functions prevent regressions from reaching production. DynamoDB is the appropriate low-latency store for agent memory across steps. CloudWatch Logs are useful diagnostically, but do not directly satisfy the listed requirements.

Answer: B

Why:
Dynamic and hierarchical chunking prevents context window overflow by selecting only the most relevant sections for augmentation. Limiting context size directly addresses truncation and prompt confusion while preserving accuracy. This approach improves reliability without significantly increasing latency and aligns with best practices for RAG troubleshooting.

Why not:

Option A: Temperature tuning affects randomness, not context overflow or truncation.

Option C: Larger context windows increase cost and do not fix inefficient retrieval strategies.

Option D: Retries do not resolve deterministic prompt construction issues and may worsen latency.

Answer: B

Why:
Dynamic and hierarchical chunking prevents context window overflow by selecting only the most relevant sections for augmentation. Limiting context size directly addresses truncation and prompt confusion while preserving accuracy. This approach improves reliability without significantly increasing latency and aligns with best practices for RAG troubleshooting.

Why not:

Option A: Temperature tuning affects randomness, not context overflow or truncation.

Option C: Larger context windows increase cost and do not fix inefficient retrieval strategies.

Option D: Retries do not resolve deterministic prompt construction issues and may worsen latency.