AIP-C01（日语） 练习题 — AIP-C01:生成式 AI 开发-专业级（日语）

This question maps to the Responsible Generative AI and Generative AI Application Monitoring domains of the AWS Certified Generative AI Developer - Professional (AIP-C01) exam. The core requirements are cross-demographic fairness measurement, real-time monitoring of fairness metrics, threshold-based alerts, weekly performance comparison of two prompt variants, and minimal custom development. The suggested solution uses Amazon SageMaker Clarify, AWS's purpose-built tool for bias detection and fairness evaluation for both traditional ML and generative AI workloads including those running on Amazon Bedrock. SageMaker Clarify natively calculates standard fairness metrics across defined demographic groups without requiring custom statistical analysis code. Publishing these metrics to CloudWatch enables real-time monitoring, and CloudWatch composite alarms can be configured to trigger alerts when the 15% discrepancy threshold between groups is breached. CloudWatch also supports scheduled weekly report generation to compare the fairness performance of the two prompt variants, eliminating the need to build custom reporting tools. This approach fully addresses all requirements with the lowest custom development overhead. Option Analysis: A. Incorrect. This approach requires developing custom Lambda functions to perform post-processing fairness analysis on model outputs, which introduces significant custom development effort that is not required with purpose-built tools. Default Amazon Bedrock API metrics do not include fairness or bias metrics, so custom logic to calculate cross-demographic recommendation discrepancy increases operational overhead, violating the least custom development requirement. B. Incorrect. Amazon Bedrock Guardrails are designed for content policy enforcement, such as filtering harmful content, redacting PII, and blocking non-compliant responses, not for measuring statistical fairness and recommendation bias across demographic groups. The InvocationsIntervened metric only tracks when guardrails modify or block a response, which does not measure discrepancies in recommendation quality or relevance between demographic cohorts, so it cannot meet the fairness evaluation requirement. C. Correct. As noted in the answer analysis, Amazon SageMaker Clarify is purpose-built for fairness and bias evaluation for generative AI workloads integrated with Amazon Bedrock. It natively computes required fairness metrics across predefined demographic groups with no custom analysis logic required. Publishing these metrics to Amazon CloudWatch enables real-time monitoring, and CloudWatch composite alarms can be configured to trigger alerts when the 15% discrepancy threshold is breached. CloudWatch also supports scheduled weekly reports to compare the performance of the two prompt variants, fully addressing all requirements with minimal custom development effort. D. Incorrect. Amazon Bedrock model evaluation jobs are designed for offline, batch evaluation of model and prompt performance, not real-time continuous monitoring of fairness metrics. Additionally, the InvocationsIntervened metric is tied to Bedrock Guardrails, which do not measure cross-demographic recommendation fairness, so this approach cannot meet the real-time alerting and fairness measurement requirements. Key Concepts: 1. Responsible Generative AI Fairness Evaluation: SageMaker Clarify provides pre-built, compliant bias detection and fairness metric calculation for generative AI workloads, eliminating the need to build custom statistical analysis logic for cross-cohort performance measurement, a core competency for the AIP-C01 exam. 2. Generative AI Monitoring Integration: AWS generative AI services including Amazon Bedrock and SageMaker integrate natively with Amazon CloudWatch for metric storage, threshold-based alerting, and scheduled reporting, reducing custom development overhead for production generative AI application monitoring. 3. Prompt Variant Performance Benchmarking: When comparing multiple prompt variants for production generative AI applications, purpose-built evaluation tools such as SageMaker Clarify reduce the overhead of building custom benchmarking and bias detection pipelines, a common use case tested in the AIP-C01 exam. References: Amazon SageMaker Clarify Fairness and Bias Detection, https://docs.aws.amazon.com/sagemaker/latest/dg/clarify-fairness-and-bias-detection.html Evaluate Generative AI Models Using Amazon Bedrock and SageMaker Clarify, https://docs.aws.amazon.com/bedrock/latest/userguide/model-evaluation-sagemaker-clarify.html

This scenario requires three distinct compliance controls for the finance company's AI assistant: blocking high-risk financial advice themes, preventing all content related to competitors, and ensuring only content aligned with the company's approved financial guidance is generated. Amazon Bedrock Guardrails provides purpose-built features to address each of these requirements without redundant or misconfigured controls. The combination of options A, D, and F directly maps to each requirement, aligns with Bedrock Guardrails best practices for regulated industry use cases, and enforces the necessary controls to avoid regulatory violations. This solution aligns with the AIP-C01 certification domain of implementing responsible generative AI controls and compliance for AWS-based generative AI applications. Option Analysis: A. Correct. Denied topics in Amazon Bedrock Guardrails are purpose-built to block entire custom conversational themes that an organization prohibits. The high-risk patterns, including requests for specific stock recommendations or guaranteed returns, are distinct thematic topics that fall under prohibited financial advice, so adding these to denied topics is the appropriate configuration to block all conversations related to these high-risk themes. B. Incorrect. Content filter guardrails in Bedrock are designed to filter predefined harmful content categories such as hate speech, violence, or sexual content based on severity levels, not custom business-specific conversational patterns. Denied topics are the correct feature for enforcing business-specific topic restrictions, so this option uses the wrong Guardrails component. C. Incorrect. Content filter guardrails are not intended for blocking specific named entities like competitor names. Custom word filters are the purpose-built Bedrock Guardrails feature for blocking specific terms and entities, so this configuration uses the wrong component for the competitor content requirement. D. Correct. Custom word filters in Amazon Bedrock Guardrails allow organizations to define specific terms, phrases, or named entities such as competitor names to block across both user inputs and model outputs. Setting input and output actions to block ensures the AI assistant never processes prompts referencing competitors or generates content about competitors, directly meeting this requirement. E. Incorrect. A low grounding score threshold would allow model outputs that have low confidence of alignment with the company's approved financial guidance reference sources. This would increase the risk of ungrounded, non-compliant claims, which violates the requirement for factually grounded content aligned with official company guidance. F. Correct. Grounding checks in Amazon Bedrock Guardrails validate that model outputs are factually consistent with organization-provided reference sources, in this case the company's approved financial guidance. A high grounding score threshold ensures only outputs with high confidence of alignment with the approved sources are allowed, blocking unsubstantiated or non-compliant claims to meet the factual accuracy requirement. Key Concepts: 1. Amazon Bedrock Guardrails Denied Topics: This feature enables organizations to define custom, business-specific conversational themes that are prohibited, with configurable block actions for both user inputs and model outputs that reference the denied themes, designed for policy enforcement unique to an organization's industry or use case. 2. Amazon Bedrock Guardrails Grounding Checks: This capability verifies that generative AI outputs are factually consistent with user-provided reference data sources, using a configurable confidence score threshold to block ungrounded or misaligned content, a critical control for regulated industries where content accuracy is legally required. 3. Amazon Bedrock Guardrails Custom Word Filters: This feature allows blocking of specific user-defined terms, phrases, or named entities across both user prompts and model outputs, ideal for enforcing restrictions on referencing third parties, restricted terminology, or prohibited language specific to an organization's policies. References: Amazon Bedrock User Guide - Guardrails for Amazon Bedrock, https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html Amazon Bedrock User Guide - Grounding checks in guardrails, https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-grounding.html

The core problem in this scenario is that synchronous RequestResponse invocations of the Bedrock RetrieveAndGenerate API force the client to wait for the full response to be generated and returned before any content is rendered, leading to hard timeouts and poor perceived performance, especially for complex queries that require longer Bedrock processing times. This use case falls under the AIP-C01 domain of optimizing generative AI application performance and user experience for interactive workloads. The correct solution leverages the existing deployed stack (Amplify, React, AppSync, Bedrock Knowledge Bases) to implement streaming responses with minimal architectural changes, which eliminates timeout errors by delivering incremental content chunks as they are generated, and improves user experience by reducing perceived latency. The suggested answer A uses the native Amplify AI Kit, which is purpose-built for this exact stack, eliminating the need for custom or rearchitected infrastructure. Option Analysis: A. Correct. The Amplify AI Kit is purpose-built for the exact stack used in the scenario: React applications, Amplify hosting, AppSync GraphQL APIs, and Amazon Bedrock services including Knowledge Bases. It provides out-of-the-box support for streaming RetrieveAndGenerate responses over GraphQL, so no major infrastructure rework is required. It also includes client-side rendering optimizations to display streamed content incrementally as it is generated, which both eliminates timeout issues from long-running complex queries and drastically improves perceived performance for end users, fully meeting all requirements. B. Incorrect. Increasing the Lambda resolver timeout and adding exponential backoff only addresses the hard timeout error, but does not resolve the poor user experience of waiting for a full response to generate for complex queries, which can still take 10+ seconds to return even with extended timeouts. This is a temporary band-aid that does not meet the requirement to enhance user experience. C. Incorrect. Adding an SQS queue between the client and AppSync introduces asynchronous processing that breaks the real-time interactive nature of an AI assistant, requiring clients to poll for results instead of receiving immediate incremental responses. This adds unnecessary custom development complexity and worsens user experience instead of improving it, and AppSync resolvers do not natively support polling SQS queues. D. Incorrect. While InvokeModelWithResponseStream supports streaming responses, this option requires replacing the existing AppSync GraphQL API with an API Gateway WebSocket API, which is a major, unnecessary rearchitecture of the existing deployed stack. The Amplify AI Kit provides native streaming support for the existing AppSync implementation, so this option introduces excessive development overhead and does not leverage existing resources optimally. Key Concepts: 1. Generative AI Response Streaming: A core AIP-C01 knowledge domain covering the practice of returning partial generative AI output chunks to the client as they are generated, rather than waiting for full completion. This pattern reduces perceived latency, avoids hard timeouts for long-running generation tasks, and is required for high-quality interactive generative AI user experiences. 2. Amplify AI Kit Native Integrations: A key AIP-C01 topic for full-stack generative AI development on AWS, the Amplify AI Kit provides prebuilt, low-code connectors for React applications using Amplify, AppSync, and Amazon Bedrock, including native support for streaming Knowledge Base responses over GraphQL without custom infrastructure changes. 3. Generative AI Invocation Pattern Selection: The AIP-C01 exam tests competence in selecting appropriate invocation patterns for generative AI workloads: synchronous RequestResponse invocations are only suited for short, low-latency tasks, while streaming invocation patterns are required for interactive, long-running generation tasks to avoid timeouts and poor user experience. References: Amplify Documentation - Generate AI content with Amazon Bedrock, https://docs.aws.amazon.com/amplify/latest/userguide/ai-gen-ai-bedrock.html AWS AppSync Documentation - Amazon Bedrock resolvers, https://docs.aws.amazon.com/appsync/latest/devguide/bedrock-resolver.html

This question aligns with the AWS Certified Generative AI Developer - Professional domain of designing scalable, compliant generative AI application architectures that integrate multiple foundation models. The scenario requires a solution that supports dynamic, code-free updates to routing rules, immediate propagation of rule changes across high concurrency workloads, and complex business logic evaluation for selecting Amazon Bedrock FMs based on compliance, cost, and performance requirements. The suggested answer C addresses all these requirements by leveraging AWS AppConfig for dynamic, no-code configuration management and Lambda as a centralized routing layer. AppConfig eliminates the need for code deployments when updating routing rules, FM IDs, cost thresholds, or compliance policies, and its Lambda agent caches configurations locally to ensure low latency even for thousands of concurrent requests. The centralized Lambda layer runs the required custom business logic to evaluate request attributes, real-time cost metrics, and compliance rules against the latest configuration to select the optimal Bedrock FM, with a single API Gateway endpoint providing a unified user-facing interface that requires no client-side changes when routing rules or FMs are updated. Option Analysis: A. Incorrect. Lambda environment variables require a full Lambda function deployment to update, which violates the requirement to switch FMs and update rules without deploying new code. Updates to environment variables also do not propagate immediately to all active Lambda execution environments, failing the immediate propagation requirement for hourly rule changes. Additionally, API Gateway cannot independently evaluate complex dynamic routing rules that rely on real-time metrics stored as static environment variables. B. Incorrect. API Gateway stage variables and request transformation templates require a re-deployment of the API stage to update, which violates the no-code change requirement for routing rule and FM switches. Request transformation templates also have limited support for complex business logic such as evaluating real-time cost metrics, A/B test splits, or dynamic regional compliance rules that change hourly, making this option unsuitable for the scenario. C. Correct. This solution meets all stated requirements. The AWS AppConfig Agent for Lambda caches routing configurations locally, enabling low-latency access for high concurrency requests while supporting immediate, code-free propagation of rule and FM configuration changes. The centralized Lambda function can run full custom business logic to evaluate all required routing factors including user tier, transaction value, regulatory zone, and real-time cost metrics to select the appropriate Amazon Bedrock FM. The single API Gateway endpoint provides a unified entry point that requires no client changes when routing rules or FMs are updated, and no code deployments are required to modify routing logic or switch between FMs. D. Incorrect. Lambda authorizers are purpose-built for authentication and authorization workflows, not for routing logic, so misusing them for routing introduces unnecessary latency and operational anti-patterns. Routing to model-specific Lambda functions also requires deploying new Lambda functions when adding or switching to new FMs, violating the no new code deployment requirement. This architecture also adds unnecessary operational overhead from managing multiple model-specific Lambda functions compared to a centralized routing layer. Key Concepts: 1. Dynamic Configuration for Generative AI Workflows: AWS AppConfig is the recommended managed service for managing dynamic routing rules, feature flags, compliance policies, and cost thresholds for generative AI applications, as it supports zero-code deployments of configuration changes, immediate propagation, and validation guardrails for high-risk updates, which is a core competency for the AIP-C01 exam. 2. Multi-FM Routing Architecture for Amazon Bedrock: Centralized routing layers that pull dynamic configuration from a managed service enable organizations to switch between FMs, run A/B tests, enforce regional compliance, and optimize cost without modifying client integrations or core application code, a key design pattern tested in the AIP-C01 exam. 3. Serverless Best Practices for Generative AI Workloads: Using Lambda with the AppConfig Agent layer provides low-latency, high-concurrency support for real-time generative AI use cases while reducing operational overhead, as the agent handles configuration caching and updates automatically without custom code. References: AWS AppConfig User Guide, https://docs.aws.amazon.com/appconfig/latest/userguide/what-is-appconfig.html Build a multi-model routing solution for Amazon Bedrock to dynamically select the optimal foundation model

The scenario outlines four core requirements: centralized management of business unit-specific prompt variants, minimal ongoing orchestration and post-processing maintenance, adjustable content moderation criteria over time, and enforcement of consistent response formatting, business unit-aligned tone, and content safety rules. The optimal solution leverages fully managed native Amazon Bedrock features to avoid building and maintaining custom infrastructure, as custom components introduce unnecessary operational overhead. The suggested answer uses integrated Bedrock capabilities that directly address all requirements out of the box, aligning with AWS best practices for low-maintenance GenAI application development covered in the AIP-C01 certification domain. Option Analysis: A. Correct. Amazon Bedrock Prompt Management is a native managed feature purpose-built to centrally store, version, and manage reusable prompt templates and business unit-specific variants, eliminating the need for custom prompt storage and selection logic. Amazon Bedrock Guardrails is a fully managed content policy tool that supports configurable category filters for hate speech and inappropriate content, sensitive term lists for blocking PHI and other restricted data, and can be updated over time to adjust moderation criteria without custom code changes. This option meets all requirements with no custom orchestration or post-processing maintenance, delivering the least operational overhead. B. Incorrect. While this option uses valid Bedrock features, it introduces unnecessary custom development overhead via an internal administration API for guardrail management. Native Bedrock Guardrails can be updated directly via the AWS Console, standard AWS APIs, or infrastructure as code without custom internal tooling, making this option higher maintenance than Option A. Audience-based threshold tuning is also not required to meet the stated scenario requirements. C. Incorrect. This option relies entirely on custom built components including business unit instruction injection in API calls, DynamoDB stored formatting rules, Step Functions orchestration for response validation, and Amazon Comprehend for post-generation content filtering. All of these custom components require ongoing development, monitoring, and maintenance, which violates the requirement to minimize orchestration and maintenance overhead. This approach is significantly more operationally intensive than using native Bedrock features. D. Incorrect. This option uses custom prompt storage in DynamoDB and two custom Lambda functions for prompt selection and content filtering. These custom components require regular code updates, dependency management, and monitoring, leading to far higher maintenance overhead than using the fully managed native Bedrock Prompt Management and Guardrails features. It does not meet the least maintenance requirement. Key Concepts: 1. Amazon Bedrock Prompt Management: A core AIP-C01 domain feature that enables centralized, versioned management of prompt templates and use case-specific variants across teams, eliminating the need for custom prompt orchestration infrastructure. 2. Amazon Bedrock Guardrails: A managed content enforcement capability for Bedrock applications that supports configurable category filtering, sensitive data protection, and topic blocking, removing the need to build custom post-processing moderation logic for GenAI workloads. 3. Low-Overhead GenAI Architecture: AWS best practice covered in the AIP-C01 exam that prioritizes integrated native service features over custom-built serverless workflows to reduce ongoing operational and maintenance overhead for GenAI applications. References: Amazon Bedrock Prompt Management Documentation, https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-management.html Amazon Bedrock Guardrails Documentation, https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html