Free and Premium Amazon Web Services AIP-C01 Dumps Questions Answers

Option A is the only choice that simultaneously addresses all three requirements: (1) higher output consistency for identical inputs, (2) sub-1-second performance, and (3) validated safety controls that block unsafe or hallucinated recommendations.

Provisioned throughput in Amazon Bedrock reserves capacity for the chosen model, which helps stabilize latency and reduces the chance of throttling or variable response times across Regions. This is important for a mobile app with strict latency goals and users distributed across multiple Regions. While provisioned throughput primarily improves performance predictability, it also reduces variability caused by contention during peak demand.

Amazon Bedrock guardrails provide validated safety controls to filter or block unsafe content. Semantic denial rules are appropriate for preventing dangerous brewing guidance (for example, excessively high temperatures) and for reducing hallucinated instructions that violate safety policies. Guardrails can be enforced consistently regardless of prompt-chain complexity, providing a uniform safety layer around the model outputs.

Amazon Bedrock Prompt Management supports controlled prompt versioning and approval workflows. By standardizing prompts, controlling changes, and ensuring the same prompt version is used for identical inputs, the company improves output stability and reduces drift caused by unmanaged prompt edits. Combined with strict configuration control (including fixed inference parameters such as temperature where appropriate), this improves repeatability and increases the likelihood of achieving the 99.5% consistency target.

Option B improves observability and experimentation but does not provide strong safety enforcement or latency stabilization. Option C improves performance through caching and tracing but does not provide validated safety controls and does not directly address cross-Region output consistency. Option D may improve retrieval but does not enforce safety controls or ensure repeatable outputs.

Therefore, Option A best meets the stability, performance, and safety requirements using AWS-native controls.

Option D is the best fit because it implements a reliable event-driven MLOps workflow that automates retraining and redeployment with clear orchestration, auditability, and production-grade error handling. The requirement is explicit: whenever a new file is uploaded to Amazon S3, the system must retrain and then redeploy the model used by a web application. A common AWS pattern is to use an S3 event notification to trigger an AWS Lambda function, which then starts a controlled workflow. In option D, Lambda serves as the event handler that reacts immediately to the S3 upload event and passes the necessary context (bucket, object key, dataset version) into an AWS Step Functions Standard state machine.

Step Functions Standard is appropriate for model retraining pipelines because training and deployment steps can be long-running and benefit from durable state, retries, and failure handling. It provides execution history, making it easier to troubleshoot why a particular retraining run failed and to prove which dataset version produced which model version. This operational visibility is critical when the dataset is “growing rapidly” and retraining is frequent.

Within the workflow, Amazon SageMaker Pipelines is the right service to run the ML lifecycle stages in a repeatable way: data processing (if needed), training, evaluation/quality checks, model registration, and deployment to an endpoint used by the application. SageMaker Pipelines is purpose-built for CI/CD-style ML, supporting automated redeployments when a new approved model artifact is produced. By calling a pipeline execution from Step Functions, the company can add governance gates (for example, only deploy if evaluation metrics meet thresholds), and can apply consistent rollback and notification steps when deployment fails.

The other options are weaker: A confuses inference with retraining and does not provide deployment orchestration. B adds unnecessary webhook complexity and describes an awkward event bus configuration. C introduces Autopilot/Data Wrangler, which may be useful but adds extra moving parts and is not required to meet the trigger-and-redeploy requirement.

Option A is correct because the problem is not total model runtime; the problem is perceived UI responsiveness. AWS documentation states that the Amazon Bedrock InvokeModelWithResponseStream API returns the model response as a stream. AWS also explains that streaming allows an application to access responses in chunks without waiting for the entire result. This lets the frontend display generated content as tokens or chunks arrive, improving the user experience even when the total inference duration remains unchanged.

This option also preserves the stated constraints. It does not require changing the foundation model, increasing token counts, provisioning extra compute, or reducing the model’s actual inference time. Instead, it changes how the response is delivered to the user. This is exactly the right pattern when users perceive slowness because the interface waits for the full completion before rendering anything. AWS examples for Bedrock streaming show applications processing response streams in real time, which supports interactive applications during high-concurrency periods.

The monitoring portion of option A is also aligned with AWS guidance. Amazon Bedrock runtime publishes CloudWatch metrics for invocation volume, latency, token consumption, error rates, and model invocation logging delivery. CloudWatch GenAI observability can track invocation count, token usage, and errors using built-in views. This supports audit and compliance visibility across Regions.

Option B adds caching of partially processed fraud explanations, which can be risky for regulated transaction-specific explanations and does not solve the UI issue for unique requests. Option C changes inputs by condensing them, which violates the constraint not to modify token counts or inputs. Option D focuses on Lambda concurrency and log retention, which does not address waiting for the full model response. Therefore, response streaming with CloudWatch and tracing is the correct solution.

Option D is the correct solution because it relies primarily on managed, purpose-built AWS services and minimizes custom infrastructure and model management. Amazon Bedrock guardrails provide native, configurable content safety controls that can block or redact disallowed content before or after model inference. This directly ensures that the platform does not process or produce inappropriate outputs while maintaining low operational overhead.

Using Amazon Comprehend PII detection as a preprocessing step integrates cleanly with an Amazon S3–based ingestion workflow. Comprehend is a fully managed service that detects and optionally redacts PII in text without requiring custom models or pipelines. This ensures that sensitive information is removed before content is passed to Amazon Bedrock for generation.

Amazon Rekognition image moderation is purpose-built for detecting unsafe or inappropriate visual content and integrates naturally into Step Functions workflows. Step Functions provides orchestration without requiring servers or long-running infrastructure, allowing the company to integrate text and image moderation steps in a clear, auditable pipeline.

Option A introduces redundant monitoring logic and alarms that do not directly enforce content safety. Option B requires building and maintaining custom SageMaker models, increasing complexity and operational burden. Option C applies moderation at authentication time and uses services like Textract that are not designed for content moderation, increasing latency and management overhead.

Therefore, Option D best satisfies content safety, PII protection, S3 integration, and minimal infrastructure management requirements.

The correct combination is A, D, and F because these guardrail features directly map to the stated financial compliance and governance requirements.

Option A is required because denied topics guardrails are explicitly designed to block entire categories of requests, such as requests for guaranteed returns or specific stock recommendations. These are regulatory-sensitive scenarios where partial filtering is insufficient and full blocking is required to prevent violations.

Option D is correct because custom word filters are the appropriate guardrail mechanism to block references to specific competitor names. Content filters are category-based (such as hate, sexual, or violence-related content) and are not suitable for blocking organization-specific competitor references. Custom word filters allow precise blocking at both input and output stages.

Option F is required because a high grounding score threshold enforces that model outputs must be strongly supported by approved source material. This prevents the AI assistant from making speculative or unfounded claims that are not aligned with the company’s approved financial guidance, which is critical in regulated financial environments.

Option B is incorrect because content filters do not target domain-specific financial advice patterns. Option C is incorrect for the same reason—competitor names are not a content filter category. Option E would weaken factual grounding and increase hallucination risk.

Therefore, A, D, and F together provide topic blocking, competitor exclusion, and factual grounding enforcement.

Option C is correct because the issue is retrieval quality: responses contain irrelevant information and sometimes miss important surrounding context. Amazon Bedrock Knowledge Bases supports hierarchical chunking with parent and child chunks. In this pattern, smaller child chunks improve precise retrieval, while larger parent chunks provide broader surrounding context when returned for generation. AWS documentation states that hierarchical chunking supports two levels: a parent chunk size and a child chunk size, with configurable overlap. This directly addresses the textbook problem because dense technical materials often need both fine-grained matching and enough context to avoid incomplete or misleading answers.

Hybrid search is also the strongest retrieval option for domain-specific textbooks. AWS documentation for Knowledge Bases explains that HYBRID search uses both vector embeddings and raw text, while SEMANTIC search uses only vector embeddings. For technical content, equations, terms, acronyms, product names, and exact phrases may be missed by pure semantic search. Hybrid search reduces this risk by combining semantic similarity with keyword matching. The Knowledge Base vector search configuration also supports controls such as number of source chunks, metadata filters, search type override, and reranking configuration, all of which help tune retrieval quality.

Adjusting the relevance threshold, or filtering low-confidence retrieval results in the application layer, helps reduce hallucinations because the model should not generate answers from weak or unrelated context. The key GenAI principle is that RAG quality depends on retrieving the right evidence before generation. If irrelevant chunks are passed to the FM, the FM can produce irrelevant or unsupported answers.

Option A uses very small fixed chunks, which can miss context across sections. Option B adds caching and feedback loops but does not fix the underlying retrieval issue. Option D’s semantic chunking can help topic boundaries, but it does not explicitly solve exact-term retrieval or low-confidence filtering. Therefore, hierarchical chunking plus hybrid search and relevance filtering is the best answer.

==========

Option C is the correct solution because it directly addresses the root cause of the problem—overly broad retrieval—while requiring minimal architectural change. Amazon Bedrock Knowledge Bases support metadata-aware filtering, which allows the system to constrain retrieval queries based on indexed metadata such as content type, publication date, source, or category.

By indexing Amazon S3 object metadata, the company can restrict vector searches to relevant subsets of the corpus, such as recent emergency reports, specific content formats, or trusted sources. This significantly reduces the number of documents evaluated during retrieval, which improves both latency and result relevance without changing embedding models or retrieval infrastructure.

This approach aligns with AWS best practices for optimizing RAG systems: when embeddings are already strong, retrieval quality is often improved by narrowing the candidate set rather than increasing model complexity. Metadata filtering reduces noise and ensures that retrieved documents are more contextually aligned with user queries.

Option A requires retraining or adapting embedding models, which the company has already determined will not provide additional benefit. Option B introduces a migration to OpenSearch, which adds operational overhead and deviates from the existing Bedrock knowledge base architecture. Option D requires moving to a different indexing service, increasing complexity and implementation effort.

Therefore, Option C provides the most effective and low-effort solution to improve retrieval accuracy and performance in the existing Amazon Bedrock RAG system.

Option D is correct because it separates the United States and European workloads and uses the correct Amazon Bedrock inference pattern for European data residency. AWS documentation states that geographic cross-Region inference keeps data processing within specified geographic boundaries such as US, EU, and APAC, and is intended for organizations with data residency and compliance requirements. The regional availability documentation further explains that a geographic profile routes within a defined geography and that prompts and outputs may move within that geography but not outside it. This directly satisfies the requirement that European customer data must be processed only in Europe-based AWS Regions.

Provisioned throughput is appropriate for the US workload because the system has a demanding and predictable performance target: 1,000 transactions per second with sub-500 ms response times. Provisioned capacity is generally used when an application needs reserved throughput and consistent performance rather than depending only on on-demand capacity. Automatic scaling based on utilization metrics supports changing demand while maintaining performance. EventBridge and Lambda can coordinate failover workflows or routing updates when connectivity disruptions occur, supporting high availability across Regions.

Option A is not the best answer because it includes a global inference profile. AWS documentation states that global cross-Region inference can route requests across supported commercial AWS Regions worldwide and has no geographic data-residency restriction. AWS explicitly notes that organizations with data residency or compliance requirements should assess whether global cross-Region inference fits their framework because requests may be processed in other supported commercial Regions. That conflicts with the European data processing requirement.

Option B uses separate Regional provisioned throughput, but it relies on a custom Application Load Balancer approach for Bedrock routing and does not use the Bedrock geographic inference profile that is designed for EU data residency. Option C uses response streaming, which improves perceived response latency for text generation but does not guarantee sub-500 ms fraud-processing latency or European-only processing. Therefore, option D is the most complete AWS-aligned solution.

Options A and E are correct because the failure pattern is caused by exceeding the model’s usable context window and by sending too much raw customer data directly to the FM. Amazon Bedrock Knowledge Bases follows the RAG pattern by splitting source content into manageable chunks, converting chunks into embeddings, storing them in a vector index, and using semantic similarity to retrieve relevant content for a user query. This supports option E because the application can index customer purchases, profiles, and review-related information, then retrieve only the most relevant records for the current recommendation request instead of stuffing all raw data into one prompt.

Option A is also correct because chunking and staged synthesis are standard ways to handle inputs that exceed an FM’s context capacity. Processing smaller segments avoids context-length-exceeded errors, and a final synthesis call can combine intermediate findings into a complete recommendation. This is especially useful when the company says it wants the system to consider all customer data, not merely truncate older or less important data. Amazon Bedrock Knowledge Bases also lets teams configure how content is chunked for storage and retrieval, which supports a structured approach rather than an oversized one-shot prompt.

Option B is not sufficient because truncating “less important” data conflicts with the requirement to consider all customer data. Option C violates the implied design direction because the company wants to use Knowledge Bases, and simply choosing a larger context model does not solve long-term growth of customer histories. Option D is technically incorrect: model parameters such as max_tokens control output generation limits, not the model’s maximum context window, and different Claude models have model-specific maximum values. You cannot use additionalModelRequestFields to bypass the FM’s context limit.

==========

Option A is the most comprehensive and architecturally aligned solution for meeting end-to-end data lineage, real-time PII filtering, and automated compliance reporting requirements in a medical GenAI system built on Amazon Bedrock. Each requirement maps directly to a managed AWS service that is purpose-built for governance, security, and compliance.

AWS Glue Data Catalog is designed to register datasets across multiple sources and maintain metadata that supports lineage tracking. By cataloging all inputs that flow into the Bedrock-based system, the organization can trace how data moves from ingestion through processing and storage, which is essential for regulatory audits in healthcare environments.

For real-time PII filtering , Amazon Bedrock Guardrails provide native PII detection and filtering during model inference. Guardrails operate inline with model invocation, ensuring sensitive information is blocked or redacted before responses are returned to users. This satisfies the requirement for real-time protection rather than post-processing analysis.

AWS CloudTrail delivers a complete audit trail of all Amazon Bedrock API calls, including InvokeModel requests and configuration changes. Storing these logs in Amazon S3 enables long-term retention and supports compliance audits. CloudTrail ensures traceability of who accessed the system, when, and how it was used.

To strengthen compliance monitoring, Amazon Macie continuously scans stored data for sensitive information and automatically classifies findings. Publishing Macie findings to Amazon CloudWatch Logs and visualizing them through dashboards enables near-real-time visibility into compliance posture and supports automated reporting workflows.

The other options fall short. Option B performs PII filtering at the application edge rather than at inference time and relies on scheduled analysis instead of real-time enforcement. Option C focuses on replication and document processing rather than inline GenAI governance. Option D uses services that are not designed for PII detection in text-based GenAI workflows and lacks native lineage tracking.

Therefore, A best fulfills all stated requirements using AWS-recommended governance and security capabilities.

Option D is the correct solution because it directly evaluates multilingual output consistency and quality in an automated, scalable, and deployment-gating workflow. Amazon Bedrock model evaluation jobs are designed to run large-scale, repeatable evaluations against defined datasets and to produce quantitative metrics that can be used as objective release criteria.

The core issue is semantic inconsistency across languages for equivalent inputs. The most reliable way to detect this is to create standardized test conversations where each language version expresses the same intent and constraints. Running those tests through the updated model and comparing results with similarity metrics (for example, semantic similarity between expected and actual answers, or between language variants) surfaces regressions that infrastructure testing cannot detect.

Bedrock evaluation jobs support running evaluations at scale and are well suited for processing large datasets quickly. By parallelizing evaluation runs across languages and conversations, the company can meet the 45-minute requirement while executing at least 15,000 conversations. Because the process is standardized, it also allows consistent baseline comparisons across releases.

Applying hallucination thresholds ensures that answers remain grounded and do not introduce fabricated details, which is particularly important when language-specific behavior shifts after a model upgrade. Integrating evaluation jobs into the CI/CD pipeline enables fully automated execution on every model or configuration update. The pipeline can enforce a hard quality gate that blocks deployment if thresholds are not met, preventing regressions from reaching production.

Option A focuses on performance and infrastructure bottlenecks, not multilingual response quality. Option B is post-deployment and too slow to prevent regressions. Option C normalizes inputs but does not measure multilingual output equivalence or provide robust, quantitative gating.

Therefore, Option D best meets the automation, scale, timing, and deployment-blocking requirements.

Option D best satisfies all functional, performance, and governance requirements while minimizing architectural complexity. The requirement explicitly states that all filtering must complete before content reaches the foundation model, which rules out asynchronous or streaming-based approaches that could delay enforcement.

Amazon Comprehend supports toxicity detection, prompt safety classification, and PII detection with entity redaction as managed capabilities. Running these filters in parallel ensures low end-to-end latency, which is essential for customer-facing communication platforms. Parallel execution avoids the cumulative latency that would be introduced by sequential pipelines.

Toxicity detection identifies offensive or abusive content early. Prompt safety classification detects requests for unethical, harmful, or manipulative advice, which directly addresses inappropriate advice solicitation requirements. PII detection with entity redaction ensures that customer privacy is preserved before data is sent to the FM, preventing sensitive information from being processed or echoed in generated responses.

Configuring thresholds allows fine-grained control over sensitivity while maintaining acceptable false-positive rates. Using CloudWatch metrics and alarms enables continuous monitoring of filtering behavior and intervention rates without adding custom routing or human review pipelines that would slow responses.

Option A lacks PII redaction. Option B introduces unnecessary model-building complexity and delayed PII checks. Option C adds sequential latency and introduces human review routing, which violates the response-time requirement.

Therefore, Option D provides the most robust, performant, and AWS-aligned layered content filtering solution.

Option B is the correct solution because it directly addresses both throughput bottlenecks and latency requirements using native Amazon Bedrock performance optimization features that are designed for real-time, high-volume generative AI workloads.

Amazon Bedrock supports cross-Region inference profiles, which allow applications to transparently route inference requests across multiple AWS Regions. During peak usage periods, traffic is automatically distributed to Regions with available capacity, reducing throttling, request queuing, and timeout risks. This approach aligns with AWS guidance for building highly available, low-latency GenAI applications that must scale elastically across geographic boundaries.

Token batching further improves efficiency by combining multiple inference requests into a single model invocation where applicable. AWS Generative AI documentation highlights batching as a key optimization technique to reduce per-request overhead, improve throughput, and better utilize model capacity. This is especially effective for lightweight, low-latency models such as Claude 3 Haiku, which are designed for fast responses and high request volumes.

Option A does not meet the requirement because purchasing provisioned throughput in a single Region creates a regional bottleneck and does not address multi-Region availability or traffic spikes beyond reserved capacity. Retries increase load and latency rather than resolving the root cause.

Option C improves application-layer scaling but does not solve model-side throughput limits. Client-side round-robin routing lacks awareness of real-time model capacity and can still send traffic to saturated Regions.

Option D is unsuitable because batch inference with asynchronous retrieval is designed for offline or non-interactive workloads. It cannot meet a strict 2-second response time requirement for an interactive AI assistant.

Therefore, Option B provides the most effective and AWS-aligned solution to achieve low latency, global scalability, and high throughput during peak usage periods.

Option B is the correct solution because AWS AppConfig is specifically designed to support dynamic configuration management with safe rollout, validation, and rollback, which are explicit requirements in the scenario.

By storing the active model provider configuration in AWS AppConfig, the company can switch between Amazon Bedrock model providers in real time without redeploying Lambda code. AppConfig supports deployment strategies such as canary releases, linear rollouts, and immediate deployments, allowing safe and controlled changes. If a configuration causes issues, AppConfig supports automatic rollback, reducing operational risk.

AWS AppConfig also supports schema validation, ensuring that configuration values such as model identifiers, provider names, or inference parameters are valid before being applied. This prevents misconfiguration from impacting production workloads.

Option A uses Parameter Store, which lacks native rollout strategies, validation, and automated rollback, making it unsuitable for safe real-time switching. Option C requires manual routing changes and code coupling, increasing operational overhead and deployment risk. Option D introduces unnecessary complexity by hosting configuration files in Amazon S3 when AppConfig already supports native hosted configurations.

Therefore, Option B provides the most robust, scalable, and low-maintenance solution for dynamic model switching in a serverless Amazon Bedrock inference architecture.

Option A best meets the requirements because it applies an AWS-aligned multi-agent pattern that cleanly separates responsibilities: a supervisor agent performs intent classification and orchestration, while specialized collaborator agents handle domain-specific tasks using the right knowledge sources. This structure is well suited for healthcare workflows where clinical questions, scheduling, and insurance processes require different policies, terminology, and data access boundaries.

The requirement for appropriate domain-specific responses is addressed by routing each user query to a department-focused collaborator agent that is grounded with its own department-specific knowledge base . Using Retrieval Augmented Generation with the correct knowledge base improves factual alignment and reduces cross-department leakage (for example, avoiding claims content in a clinical answer). It also supports better prompt grounding and more consistent tone and constraints per department.

The requirement to isolate data maps to using separate knowledge bases per agent and enforcing access through IAM controls , ensuring that each agent can retrieve only from the authorized datasets. This is important for minimizing unintended exposure of sensitive or irrelevant departmental data and supports governance and compliance needs.

For scalability and thousands of parallel interactions , this architecture minimizes contention and bottlenecks. Each collaborator agent can scale independently because requests are distributed across multiple agents and multiple retrieval backends. Operationally, onboarding new features is also simpler: the company can add a new collaborator agent (for example, “billing disputes” or “pharmacy refills”) with its own knowledge base and policies without redesigning the entire assistant.

Option B introduces unnecessary complexity with multiple supervisors and manual handoffs. Option C overloads a single agent with broad instructions and rule-based routing, which increases prompt complexity and reduces maintainability as features grow. Option D creates high operational complexity and risks inconsistent outputs when merging responses from parallel supervisors, and it weakens data isolation by using a shared knowledge base across agents.

The correct combination is C and D because together they enforce centralized governance over both model access and prompt content controls , which are the two core requirements of the scenario.

To ensure employees can use only approved Amazon Bedrock models , governance must be enforced at the organization level and not rely on individual application logic. Service Control Policies (SCPs) are the strongest control mechanism available in AWS Organizations because they define the maximum permissions an account or principal can have. In option C , the SCP prevents any Amazon Bedrock model invocation unless a centrally approved guardrail identifier is specified. This ensures that guardrails are always enforced, regardless of how or where the invocation originates. The additional use of IAM permissions boundaries ensures that even within allowed accounts, employees are restricted to invoking only explicitly approved foundation models.

To prevent specific topics and proprietary information from being included in prompts , Amazon Bedrock Guardrails must be used. Guardrails operate inline during model invocation and can block disallowed content before it is processed by the model. Option D correctly specifies a block filtering policy , which is appropriate when content must be prevented entirely rather than partially redacted. Deploying the guardrail using AWS CloudFormation StackSets allows the company to centrally manage and consistently deploy the same guardrail configuration across all accounts in the organization, ensuring uniform enforcement.

Option E uses mask filtering, which is better suited for redacting sensitive output rather than preventing prohibited content from being submitted in prompts. Option B attempts to use SCPs alone but does not enforce guardrail deployment or content filtering. Option A incorrectly places guardrail enforcement in permissions boundaries, which are not designed to validate request parameters such as guardrail identifiers.

By combining SCP-based enforcement with centrally deployed Bedrock guardrails , options C and D together provide strong, scalable, and centrally managed controls for both content safety and model governance across the organization.

Option A is the correct solution because it satisfies authentication, private connectivity, fine-grained authorization, and auditing using AWS-recommended patterns.

SAML federation between Microsoft Entra ID and IAM is a mature, well-supported integration that enables centralized enterprise authentication. Department-specific IAM roles allow precise control over which Bedrock ModelId values each department can invoke, enforcing access by model family.

Using AWS PrivateLink interface VPC endpoints for Amazon Bedrock runtime services ensures that all inference traffic stays on private AWS network paths, with no public internet exposure. NAT gateways and public endpoints, as used in other options, violate this requirement.

AWS CloudTrail provides authoritative audit logs of all Bedrock API calls, which is required for compliance. Amazon Bedrock model invocation logging complements CloudTrail by capturing detailed prompt and response metadata for deeper auditing and investigation.

Option B uses public endpoints via NAT. Option C incorrectly claims public endpoints can be private. Option D relies on IdP-side logs, which do not capture Bedrock API activity.

Therefore, Option A is the only solution that fully meets security, compliance, and observability requirements.

Option B meets the requirement to combine scalable automated evaluation with targeted human oversight using managed AWS GenAI capabilities. Amazon Bedrock evaluations enable systematic, repeatable quality assessment across large volumes of interactions. Using an LLM-as-a-judge approach with a strong evaluator model such as Anthropic Claude Sonnet allows the company to automatically score outputs for dimensions like factual accuracy, conversational appropriateness, and policy alignment. This directly supports “automated quality evaluations at scale” without building custom scoring models.

However, financial recommendations add higher risk because regulatory compliance requires additional enforcement beyond general quality scoring. Amazon Bedrock guardrails provide a dedicated policy enforcement layer that can block or intervene when responses violate compliance constraints. Guardrails are particularly important for preventing disallowed financial guidance patterns and ensuring consistent behavior across deployments.

The requirement also calls for “targeted human reviews of critical interactions.” Amazon Augmented AI (A2I) is a managed human review service that supports routing specific items to human reviewers based on rules or confidence thresholds. In this design, the system can automatically send only high-risk or policy-flagged interactions to qualified financial experts for review, keeping human effort focused where it matters most while maintaining scale.

Option A is not scalable because it requires manual review of all responses. Option C relies on static rules and end-user feedback, which is insufficient for regulatory compliance and factual accuracy assurance. Option D provides monitoring but not structured quality evaluation or policy enforcement.

Therefore, Option B provides the most complete, AWS-aligned solution for scalable evaluation plus human oversight in a regulated financial context.

Option A is the best solution because it directly addresses both observed problems: user-perceived latency and resolver timeouts that occur more frequently for complex prompts. In the current design, an AWS AppSync Lambda resolver is configured with synchronous RequestResponse behavior. That means the client receives nothing until the entire retrieval and generation workflow completes. For longer-running knowledge base queries, this increases the likelihood of hitting request time limits in the synchronous path and creates a poor user experience because the UI appears stalled.

Using AWS Amplify AI Kit to implement streaming responses allows the application to return partial output incrementally as the model produces tokens. This improves perceived responsiveness because users can see the answer forming immediately, even when the full response takes longer. Streaming also reduces the impact of variable model latency and retrieval time because the client no longer waits for a single final payload before rendering content. From a troubleshooting perspective, streaming makes it easier to distinguish “slow generation” from “no response,” and it provides faster feedback during testing of complex questions.

Option B is not sufficient because increasing timeouts and adding retries can worsen load and cost while still producing a stalled UI experience. Retries also risk duplicating requests to the knowledge base and can amplify token usage. Option C introduces an awkward polling model for GraphQL interactions and adds significant operational complexity, while not inherently improving interactivity. Option D adds major architectural changes by replacing the knowledge base RetrieveAndGenerate call path with a different streaming invocation API and introducing a WebSocket layer, which is unnecessary when the goal is primarily to fix timeouts and improve UX within the existing AppSync and Amplify design.

Therefore, streaming through Amplify AI Kit is the most effective and lowest-friction improvement.

Thought for 24s

The correct combination is B and C because together they address both workload diversity and hybrid deployment requirements with minimal custom engineering.

Option B provides consistent, high-throughput access by configuring provisioned throughput in Amazon Bedrock. Provisioned throughput guarantees predictable capacity and performance, which is essential for batch processing workloads that require sustained inference rates. This eliminates cold starts and throttling concerns that can occur with purely on-demand usage, making it well suited for high-volume enterprise workloads.

Option C enables hybrid deployment across cloud and on-premises environments by deploying foundation models to Amazon SageMaker AI endpoints and using Amazon SageMaker Neo for edge and on-premises optimization. SageMaker Neo compiles models for target hardware, allowing inference to run efficiently outside the AWS cloud while still using AWS-managed tooling. Orchestrating these deployments with AWS Lambda allows consistent invocation patterns across environments.

Option A uses asynchronous endpoints, which are not suitable for real-time, low-latency inference. Option D addresses scaling but does not support on-premises or hybrid deployment. Option E simplifies model onboarding but does not address hybrid execution or guaranteed throughput.

Therefore, Options B and C together provide real-time and batch support, predictable performance, and true hybrid deployment while minimizing operational overhead.

Option C best meets the requirements by combining native Amazon Bedrock logging with adaptive monitoring and minimal operational overhead. Amazon Bedrock model invocation logging can be sent directly to CloudWatch Logs, where detailed fields such as InputTokenCount, OutputTokenCount, and tool invocation metadata are captured for each request.

CloudWatch metric filters allow extraction of structured metrics from logs, including tool-specific token consumption patterns. By defining filters per tool integration, the company can isolate which tools are responsible for increased token usage without building custom log-processing pipelines.

CloudWatch anomaly detection provides automatic baseline modeling and dynamic thresholds based on historical traffic patterns. Unlike static alarms, anomaly detection adapts as usage evolves, making it ideal for applications with changing workloads or seasonal usage patterns. This directly satisfies the requirement to automatically adjust thresholds as traffic patterns change.

When abnormal token consumption occurs, anomaly detection alarms trigger immediately, enabling rapid investigation and remediation. Because this solution uses fully managed AWS services without custom analytics jobs or manual threshold tuning, it significantly reduces operational effort.

Option A fails to adapt to changing patterns. Option B introduces batch analysis and delayed insights. Option D requires manual intervention and custom code, increasing maintenance burden.

Therefore, Option C provides the most scalable, adaptive, and low-maintenance solution for monitoring and controlling token consumption in Amazon Bedrock–based applications.

The correct combination is A, B, and D because these three options collectively satisfy the mandatory requirements for structured extraction, PII redaction before inference, regional human review, data residency, auditability, and high-scale availability with managed AWS services.

Option A is essential because Amazon Textract is the AWS-managed service designed to extract structured data from scanned documents such as forms, tables, and financial statements. Textract provides confidence scores, and Amazon Augmented AI (A2I) is purpose-built to route low-confidence extractions to human reviewers. Deploying Textract and A2I within the same Region ensures that the human review loop remains regionally constrained, meeting strict data residency requirements for applicants.

Option B satisfies the requirement to redact PII before inference by using AWS Lambda preprocessing. It also adds Amazon Bedrock guardrails to enforce safety controls on model outputs. Region-specific IAM roles ensure that only authorized principals in the correct Region can access the extracted data and invoke downstream services, strengthening residency enforcement and auditability.

Option D ensures that source documents are stored in Amazon S3 in the same Region as the applicant. Object metadata and tagging provide an auditable trail, supporting compliance reporting and traceability. S3 also provides the durability and availability needed to support 99.9% application availability as part of a well-architected pipeline.

Option C is not the correct approach for structured extraction from scans. Option E adds useful quality validation but is not strictly required to meet the stated requirements compared to A, B, and D. Option F is unrelated to the extraction/redaction/residency workflow requirements.

Therefore, A, B, and D are the best three choices to meet all stated requirements with minimal operational overhead.

Option C is the correct solution because it provides a secure, scalable, and standards-compliant way to expose an MCP server to an AI agent while enforcing strong user authorization. The Model Context Protocol supports HTTP-based transports for remote MCP servers, making Streamable HTTP the appropriate choice when the server is hosted as a managed service rather than a local process.

Hosting the MCP server in AWS Lambda enables automatic scaling and cost-efficient execution. By placing Amazon API Gateway in front of the Lambda function, the company creates a secure, managed HTTP endpoint that the AI agent can invoke reliably. This architecture cleanly separates transport, authentication, and business logic, which aligns with AWS serverless best practices.

Using Amazon Cognito to enforce OAuth 2.1 ensures that only authenticated and authorized users can access the MCP server. This satisfies security and compliance requirements when the MCP server handles sensitive user information. Cognito integrates natively with API Gateway, removing the need for custom authentication logic and reducing operational overhead.

Option A lacks user-level authorization controls. Option B and Option D rely on STDIO transport, which is intended for local or tightly coupled processes and is not suitable for distributed, serverless architectures. Option D also introduces security risks by handling credentials through environment variables.

Therefore, Option C best meets the requirements for secure access control, scalability, and correct MCP integration in an AWS-based AI agent architecture.

Option A is the most complete solution because it provides a fully automated canary strategy with staged traffic shifts, metric-based decisioning, and automatic rollback, all using managed AWS services. The requirement emphasizes automation, health-based traffic progression, and zero manual intervention to revert if performance degrades.

AWS Step Functions is well suited for orchestrating controlled deployment workflows with deterministic stages, waits, and conditional branches. By shifting traffic in stages and pausing for observation windows, the system can evaluate real-time inference latency and error rates before promoting more traffic to the new model version. Amazon CloudWatch provides the necessary real-time metrics and alarms for latency and error monitoring.

Invoking a Lambda function to evaluate CloudWatch metrics enables dynamic logic: increase traffic if thresholds remain healthy, reduce traffic or roll back if error rates rise or latency exceeds limits. Step Functions can halt the deployment by stopping progression or triggering rollback steps immediately, meeting the requirement for automated revert without human action.

Amazon EventBridge provides reliable automation triggers when a new model version is released, ensuring the deployment process is event-driven and repeatable.

Option B depends on “external logic,” which introduces operational risk and does not guarantee automatic rollback without custom systems. Option C incorrectly uses SageMaker endpoint variants to represent Bedrock model versions, which is not the intended integration model. Option D is overly indirect and operationally complex, using log pipelines and automation runbooks instead of direct metric-based traffic control.

Therefore, Option A best meets the requirements for automated gradual traffic shifting, real-time monitoring, and automatic rollback for Amazon Bedrock model deployments in a canary strategy.

Option B is the best solution because hierarchical chunking is specifically designed to preserve broader semantic context while still enabling precise retrieval at paragraph or sub-paragraph granularity. The problem described—answers citing irrelevant sections when a query spans multiple paper sections—often occurs when chunks are either too small (losing cross-paragraph context) or too “flat” (retrieving isolated snippets without their surrounding rationale).

In a scientific paper, related information is frequently distributed across methodology, results, and discussion. Flat, fixed-size chunking (Option A) can split these logically connected ideas into separate chunks, causing retrieval to surface fragments that match a term but not the full intent. Semantic chunking (Option C) improves boundary placement, but it does not inherently provide a multi-resolution structure that helps preserve section-level continuity at massive scale.

Hierarchical chunking solves this by creating parent chunks (larger context windows) that capture broader section context and child chunks (smaller units) that retain retrieval precision. When the retriever identifies relevant child chunks, it can also bring in the associated parent context so the foundation model sees the surrounding methodological or discussion framing. The defined overlaps further reduce the risk that key transitions or references are split across chunks.

This approach is well suited for a corpus of 25 million papers because it improves relevance without requiring a custom reranking model or a manual preprocessing pipeline. It remains operationally efficient because it is configured at the knowledge base level rather than implemented through custom code per document.

Option D introduces high operational complexity and inconsistent document handling at scale. Therefore, Option B best meets the requirement to preserve semantic context across related paragraphs and improve citation relevance across scientific paper sections.

Option B best satisfies the requirements with the least custom development effort by using native Amazon Bedrock capabilities for prompt experimentation, traffic management, fairness monitoring, and alerting. Amazon Bedrock Prompt Management allows teams to define and manage multiple prompt variants without code changes, making it ideal for comparing recommendation strategies across demographic groups.

Amazon Bedrock Flows enables controlled traffic allocation between prompt variants, which supports real-time A/B testing. This allows the company to collect live fairness metrics under production conditions instead of relying on offline analysis. Because Flows are fully managed, they eliminate the need for custom routing or experimentation frameworks.

Amazon Bedrock guardrails provide built-in monitoring and intervention mechanisms. When configured for fairness-related checks, guardrails can detect policy violations and surface metrics such as InvocationsIntervened, which indicate when outputs are modified or blocked due to rule enforcement. These metrics integrate directly with Amazon CloudWatch, enabling real-time dashboards and threshold-based alarms. Setting an alarm at a 15% discrepancy threshold satisfies the alerting requirement with minimal configuration.

Weekly reporting can be generated from CloudWatch metrics using scheduled exports or dashboards without building custom analytics pipelines. Option A requires significant custom post-processing logic. Option C introduces an additional service with higher operational overhead and is not optimized for real-time monitoring. Option D focuses on offline evaluation jobs and does not provide continuous real-time fairness monitoring.

Therefore, Option B provides the most AWS-native, scalable, and low-effort solution for fairness evaluation and monitoring.

Option A is the correct solution because hybrid search directly addresses the core retrieval failure modes while maintaining low latency and minimal operational overhead. In medical and scientific domains, exact terminology, abbreviations, and acronyms (for example, drug names, procedures, or conditions) are critical. Pure vector similarity search often underweights these exact matches, leading to missed results and excessive semantically related but irrelevant documents.

Amazon OpenSearch Service natively supports hybrid search, which combines keyword-based retrieval (such as BM25) with vector similarity search. Keyword search ensures precise matching for exact terms and acronyms, while vector search captures semantic meaning and contextual similarity. By blending these approaches, the retrieval system improves both precision and recall without introducing additional infrastructure.

Hybrid search operates within the same OpenSearch index and query path, which preserves low end-user latency even at large scale. This is especially important as the document collection grows to millions of documents. Because OpenSearch handles scoring and ranking internally, no additional orchestration layers or post-processing steps are required.

Option B increases computational cost and latency while failing to address exact-term recall. Option C introduces a new service and ingestion pipeline, increasing operational overhead and latency. Option D adds model hosting, re-ranking infrastructure, and complexity that is unnecessary when OpenSearch provides native hybrid retrieval.

Therefore, Option A delivers the best balance of retrieval quality, scalability, latency, and operational simplicity for medical RAG workloads.

Option B is the correct solution because it provides a scalable, durable, and secure architecture for conversational GenAI workloads that require multi-step clarification workflows and persistent memory.

AWS Step Functions Standard workflows are designed for long-running, stateful workflows with high reliability, which is ideal for clarification loops that may require multiple back-and-forth interactions. The Wait for a Callback pattern allows the workflow to pause while awaiting additional user input, making it well-suited for handling ambiguous queries without losing execution state.

Storing conversation history in Amazon DynamoDB enables millisecond-latency reads and writes at massive scale, supporting thousands of concurrent users. DynamoDB’s on-demand capacity mode automatically scales with traffic, eliminating capacity planning. Server-side encryption ensures that stored conversation data is encrypted at rest, meeting security and compliance requirements for personalized data.

Option A uses Step Functions Express and Amazon RDS, which is not ideal for long-lived conversational workflows and introduces scaling and connection management challenges. Option C stores conversations as individual S3 objects, which increases latency and complicates context retrieval. Option D relies on Amazon ElastiCache, which is optimized for ephemeral caching rather than durable, auditable conversation history.

Therefore, Option B best balances scalability, performance, durability, and security for a conversational Amazon Bedrock–based customer support application.

Option B is the best solution because it directly addresses the Step Functions 256 KB state payload quota by externalizing large intermediate artifacts to Amazon S3 and passing only lightweight references (URIs/keys) between states. This is a standard AWS pattern for workflows that produce large intermediate results, and it avoids introducing additional databases, compression logic, or cross-state-machine coordination that increases operational overhead.

In a multi-agent ReAct workflow, intermediate reasoning traces can be verbose and grow quickly as each agent produces chain-of-thought style artifacts, structured outputs, and supporting evidence. Step Functions is designed to orchestrate state transitions and pass JSON payloads, but large payloads should be stored outside the state machine and referenced by pointer values. Using Amazon S3 for intermediate outputs is operationally efficient because the application already uses S3 for storage, and S3 provides durable, low-cost storage with simple access patterns.

ResultPath and ResultSelector allow each state to store or reshape results so that only the required reference fields (such as s3Uri, object key, metadata, trace IDs) are forwarded to subsequent states. This preserves observability because the workflow can still log trace references, correlate steps with S3 objects, and store structured metadata for debugging. It also preserves the sequential validation design, keeping the existing ReAct pattern intact while preventing failures due to oversized payloads.

Option A adds additional services and read/write patterns that increase operational complexity. Option C introduces custom compression/decompression logic that is fragile, adds latency, and complicates troubleshooting. Option D increases orchestration overhead by splitting workflows and coordinating with events, which makes debugging harder and increases failure modes.

Therefore, Option B meets the payload limit requirement while keeping the architecture simple and observable.

Option B best satisfies the requirements because it directly applies Retrieval Augmented Generation principles using managed Amazon Bedrock Knowledge Bases, which are designed to handle large, complex documents while preserving contextual relationships. Financial reports often interleave tables with explanatory narrative, and accurate analysis depends on keeping those elements logically connected. By segmenting documents based on their structural layout —for example, sections, subsections, tables, and surrounding commentary—the knowledge base can retrieve semantically relevant chunks that maintain this relationship during inference.

Amazon Bedrock Knowledge Bases support contextual chunking strategies that go beyond simple fixed-size segmentation. This is critical for financial documents, where a metric in a table may be explained in adjacent paragraphs or footnotes. Context-aware chunking ensures that retrieved content includes both the numeric data and its interpretation, enabling the foundation model to generate accurate, grounded responses. Including citations further improves analyst trust and auditability by allowing users to trace answers back to specific source sections, which is a common requirement in financial environments.

Scalability is another key requirement. Knowledge Bases manage embedding generation, indexing, and retrieval orchestration as a managed service, which allows the solution to scale across large document collections without requiring custom infrastructure or model hosting. This approach also supports efficient updates as new quarterly reports are added, ensuring the retrieval layer remains current.

Option A does not scale well because processing entire 5–100 page documents in a single prompt increases token usage, latency, and cost while risking context truncation. Option C relies on fixed-size chunking triggered at query time, which often breaks semantic relationships in structured financial content. Option D introduces unnecessary architectural complexity by splitting structured and unstructured data into separate applications, increasing operational overhead without providing better contextual retrieval than a unified RAG approach.

Option C is the correct solution because it directly addresses both correctness and performance issues by grounding the model’s responses in authoritative product data using Retrieval Augmented Generation. Amazon Bedrock Knowledge Bases are designed to connect foundation models to trusted enterprise data sources, ensuring that generated responses are constrained to known, validated content.

By ingesting the product catalog into a knowledge base, the GenAI application retrieves only products that actually exist in the catalog. This prevents hallucinated or unavailable recommendations, which is a common issue when models rely solely on prompt instructions without retrieval grounding. RAG ensures that the model’s output is based on retrieved facts rather than learned generalizations.

Setting the PerformanceConfigLatency parameter to optimized enables Bedrock to prioritize lower-latency retrieval and inference paths, improving responsiveness for real-time recommendation scenarios. This directly addresses the reported performance issues without requiring provisioned throughput or caching strategies that are ineffective for mostly unique interactions.

Option A improves safety and latency predictability but does not ensure recommendations are limited to valid products. Option B relies on prompt constraints, which are not sufficient to prevent hallucinations. Option D introduces additional validation and caching layers but increases complexity and does not improve generation relevance.

Therefore, Option C best resolves both relevance and latency challenges using AWS-native, low-maintenance GenAI integration patterns.

Option A provides the most complete, AWS-native defense-in-depth solution for protecting against prompt injection attacks while meeting audit and resiliency requirements. Amazon Bedrock guardrails are designed specifically to enforce safety policies on both user inputs and model outputs, including protections against prompt injection and jailbreak attempts.

Setting content filters to high increases sensitivity to malicious or manipulative inputs. Guardrail profiles allow the same guardrail configuration to be applied consistently across multiple Regions, enabling cross-Region inference and failover without configuration drift. This directly satisfies the requirement for regional resilience.

Amazon CloudWatch Logs captures detailed guardrail intervention events, including when content is blocked, modified, or flagged. Custom metrics derived from these logs enable fine-grained auditing, alerting, and reporting on safety enforcement actions. This provides a more detailed audit trail of safety interventions than API-level logs alone.

Option B adds WAF protection but lacks detailed guardrail intervention logging. Option C introduces additional services and custom logic that increase complexity and may miss model-specific injection patterns. Option D references replication concepts that are not aligned with Bedrock guardrail operational models and relies on word filters, which are insufficient against sophisticated prompt injection techniques.

Therefore, Option A best meets the requirements for layered protection, auditability, and cross-Region resilience using managed Amazon Bedrock safety controls.

The correct answers are A and D because they directly reduce time-to-first-token and stabilize p95 latency for interactive, real-time chat workloads hosted on Amazon SageMaker AI real-time endpoints.

Option D addresses the biggest driver of uneven latency: cold starts and scale-to-zero behavior. By setting the minimum number of instances to greater than 0, the endpoint always has warm capacity and loaded runtime resources, eliminating the first-request penalty that causes users to wait multiple seconds. Enabling response streaming improves perceived latency by returning the first tokens as soon as they are generated rather than waiting for the complete response. This directly targets the abandonment problem described (users leaving after waiting for the first token).

Option A further improves p95 latency and throughput by removing model loading overhead during inference and improving GPU utilization. Preloading model weights during container startup ensures the model is ready before traffic arrives and avoids unpredictable on-demand weight loading. Dynamic batching increases efficiency by grouping compatible requests into a single inference pass, reducing per-request overhead and improving GPU saturation. When tuned properly for interactive workloads, batching can reduce tail latency while preserving responsiveness by enforcing small batch windows.

Option B makes latency worse because setting minimum instances to 0 and lazily loading weights guarantees cold-start delays and unpredictable first-token performance. Option C similarly increases cold-start behavior through lazy loading and offers no batching benefits. Option E is designed for non-interactive workloads and introduces queueing and storage latency, which conflicts with the 800 ms p95 requirement for interactive chat.

Therefore, A and D are the best combination to achieve consistently low p95 latency and fast first-token streaming for a SageMaker-hosted chat assistant.

Option C best satisfies the requirement to change routing decisions without redeploying code while supporting complex, frequently changing business logic at scale. AWS AppConfig is designed for centrally managing dynamic configuration (feature flags, rules, thresholds, and policy parameters) and deploying changes safely. It supports controlled deployments, validation, and rapid propagation of updated configuration values, which aligns with “real-time cost metrics that change hourly” and the need for “immediate propagation across thousands of concurrent requests.”

In this design, the Lambda function becomes the policy decision point. For each request, it evaluates user attributes (tier, transaction value), context (regulatory zone, Region), and live cost/performance thresholds stored in AppConfig to determine which Amazon Bedrock FM to invoke. Because the routing rules and FM identifiers are delivered as configuration, the company can switch models, adjust A/B testing weights, or update compliance routing rules by deploying new AppConfig configuration versions rather than pushing new application code. This reduces operational risk and accelerates iteration.

Exposing a single API Gateway endpoint also minimizes client complexity and keeps routing logic server-side, which is important when rules change frequently. Lambda can cache configuration between invocations (within the execution environment) to reduce repeated fetch overhead while still picking up changes quickly, enabling both low latency and rapid rule rollout under high concurrency.

Option A relies on Lambda environment variables, which are not intended for frequent real-time updates and typically require function configuration updates that are slower and operationally brittle. Option B uses mapping templates and stage variables, which are limited for complex rule evaluation and safe rollout patterns. Option D misuses authorizers for business routing, adds extra latency and complexity, and complicates observability and error handling by splitting decisioning from execution.

Option A best meets the combined requirements of low latency, stability, and validated safety controls by using purpose-built Amazon Bedrock features designed for production GenAI operations. The company’s latency target of under 1 second and its observation of degradation during spikes strongly indicate capacity and throughput variability. Provisioned throughput for Amazon Bedrock is intended to deliver more predictable performance by reserving inference capacity for a chosen model, reducing throttling risk and stabilizing response times under load. This directly improves operational consistency across Regions where on-demand capacity can vary.

The requirement to “block unsafe or hallucinated recommendations” is most directly addressed by Amazon Bedrock Guardrails . Guardrails provide managed safety enforcement, including sensitive information controls and configurable content policies. Using semantic denial rules enables the application to prevent unsafe guidance such as dangerous brewing temperatures or other harmful procedural instructions, enforcing safety at the model boundary rather than relying on downstream filtering.

The remaining requirement is “99.5% output consistency for identical inputs.” While generative models can be probabilistic, production systems achieve practical consistency by controlling prompt versions, inputs, and policy behavior. Amazon Bedrock Prompt Management supports controlled prompt lifecycle practices, including versioning and approval workflows, which reduce unintended drift across deployments and Regions. By ensuring the same approved prompt templates and parameters are used consistently, the company can materially improve repeatability for the same structured inputs and retrieval context, which is essential in multi-stage prompt chains.

The other options are incomplete. B improves experimentation and observability but does not enforce safety controls or stabilize latency. C can improve performance, but it does not provide validated safety enforcement at inference time. D can help retrieval relevance, but it does not address unsafe outputs or inference stability. Therefore, A is the only option that simultaneously targets predictable latency, governance of prompt behavior, and strong safety controls within Amazon Bedrock.

Option B is the correct solution because it uses native AWS governance, access control, and auditing capabilities to protect PII while enabling controlled FM access to authorized data subsets. AWS Lake Formation is designed specifically to manage fine-grained permissions for data lakes, including column-level access control, which is critical when handling sensitive financial and PII data.

LF-Tags allow data administrators to define scalable, attribute-based access control policies. By tagging databases, tables, and columns with business unit and Region metadata, the company can enforce policies that ensure the foundation model only accesses approved datasets with PII-redacted columns. This eliminates the risk of sensitive data leaking into production inference workflows.

IAM role-based authentication ensures that the FM accesses data using least-privilege credentials. This integrates cleanly with Amazon Bedrock, which supports IAM-based authorization for service-to-service access. AWS CloudTrail provides immutable audit logs for all access attempts, satisfying compliance and regulatory requirements.

Option A introduces unnecessary data duplication and weak governance controls. Option C relies on custom application logic, increasing operational risk and complexity. Option D bypasses Lake Formation’s fine-grained controls and relies on presigned URLs, which reduces governance visibility and control.

Therefore, Option B best meets the requirements for security, compliance, scalability, and auditability when integrating Amazon Bedrock with a Lake Formation–governed data lake.

Option A is the best solution because it satisfies streaming, token control, and retry requirements while keeping operational overhead low by using fully managed, serverless AWS services. Amazon API Gateway HTTP APIs provide a lightweight, cost-effective front door for APIs and integrate cleanly with AWS Lambda for request processing and security controls.

AWS Lambda response streaming allows the API to begin returning content to the client as soon as partial model output is available, reducing perceived latency and improving user experience for long responses. Using Lambda as the integration layer also provides a centralized place to enforce token-aware request handling, such as rejecting oversized requests, truncating optional context, or applying consistent limits across users and tenants to manage compute usage.

Retry logic is best handled in the client or integration layer for transient failures such as timeouts and throttling. Lambda can implement controlled retries with exponential backoff and jitter, while API Gateway timeouts help bound request lifetimes and prevent hung connections from consuming resources indefinitely. Because the model service is managed, the company avoids infrastructure management and focuses only on request shaping, safety, and resiliency behavior.

Option B is not suitable because client-side polling is not true streaming, front-end token enforcement is insecure and inconsistent, and API Gateway does not provide model-aware retry behavior on its own. Option C introduces container hosting and scaling complexity, which increases operational overhead compared to serverless. Option D can work, but REST APIs are generally heavier than HTTP APIs for this pattern and do not reduce overhead compared to Option A.

Therefore, Option A provides the required streaming and resiliency capabilities with the least infrastructure management effort.

Option B is correct because it uses the managed evaluation capability in Amazon Bedrock that is intended specifically for comparing foundation models using a consistent prompt set and producing structured results with minimal custom tooling. In a Bedrock evaluation workflow, you provide an input dataset of prompts, typically in JSON Lines format so each line represents one evaluation record. Storing the JSONL file in Amazon S3 allows Bedrock to read the dataset at scale and write standardized evaluation outputs back to S3 for downstream analysis, sharing, and retention.

The key requirement is to assess both quality and safety across multiple models. A Bedrock evaluation job can use a judge model to score the generated outputs against defined criteria. This approach supports repeatable, apples-to-apples comparisons because the same judge model and scoring rubric can be applied to every candidate foundation model. The candidate models are configured as generators, meaning each evaluation job run uses one selected FM to produce answers for the same prompt set, and the judge model evaluates those answers. That matches the requirement to generate evaluation reports that help a data scientist select the best FM.

Option A does not use Bedrock evaluation jobs, and a knowledge base plus RetrieveAndGenerate is a RAG pattern, not an evaluation framework. It would produce responses but not standardized scoring and reporting suitable for model selection. Option C is incorrect because Bedrock evaluation outputs are delivered to S3, not directly to a BI destination, and selecting the candidate FM as the evaluator conflicts with the intended pattern of using a stable judge model. Option D misuses knowledge bases and retrieval evaluation types when the requirement is prompt-based model assessment rather than evaluating retrieval quality.