AIP-C01 AWS Certified Generative AI Developer - Professional Questions and Answers

Option D is the best solution because it delivers a fully managed, scalable pipeline with minimal infrastructure management while meeting the 50 GB and 4-hour constraint. AWS Step Functions provides a serverless orchestration layer that can coordinate parallel processing steps, retries, and error handling without managing clusters or tuning long-running compute.

Using Amazon Comprehend for PII detection fulfills the requirement to remove customer PII in a managed and consistent way. Step Functions can coordinate Comprehend calls at scale and route sanitized outputs into the embedding step. Generating embeddings with Amazon Bedrock keeps the entire workflow within AWS managed services, eliminates the need to maintain custom embedding models, and supports consistent vector representations for downstream retrieval.

Direct integration with Amazon OpenSearch Serverless provides a low-operations vector store that can handle large-scale indexing and similarity search without cluster sizing, node maintenance, or shard management. This aligns strongly with the requirement for least operational overhead and supports growth beyond the initial 50 GB corpus. Step Functions can batch and parallelize ingestion into OpenSearch Serverless to meet the 4-hour completion goal in a cost-effective manner by controlling concurrency, chunk sizes, and failure handling.

Option A can be difficult and costly at this scale because Lambda concurrency and per-invocation overhead can become complex to tune for 50 GB within 4 hours. Option B introduces SageMaker Processing and embedding model management, increasing operational complexity. Option C requires EMR cluster management and tuning, which is the opposite of minimal overhead.

Therefore, Option D is the most operationally efficient, scalable, and managed approach to build the required PII-sanitized embedding pipeline for a RAG corpus.

Option D requires the least development effort because it uses a managed retrieval workflow that bundles the most time-consuming parts of semantic search: embedding generation, vector indexing, and natural language retrieval. With an Amazon Bedrock knowledge base, the application does not need to implement and operate separate services to (1) generate embeddings for hundreds of millions of records, (2) store and manage vectors, (3) build query-time embedding conversion logic, and (4) implement k-NN search orchestration. Instead, the knowledge base is configured to automatically create embeddings during ingestion, and the application queries it using the Amazon Bedrock Retrieve API , which accepts natural language input and performs the vector search as a managed capability.

The performance requirement (95% of queries within 500 ms) is best served by a purpose-built vector search backend rather than running similarity search directly inside a transactional PostgreSQL system at this scale. A knowledge base is designed for retrieval patterns and can be backed by scalable vector stores, which helps meet latency goals under heavy concurrency. The hourly freshness requirement maps naturally to ingestion updates: the pipeline can re-ingest updated restaurant details on a schedule so the knowledge base remains current without building custom re-embedding workflows in application code.

Cost-effective scaling during peak periods is also easier with a managed retrieval layer because scaling the retrieval workload is separated from the operational database. This avoids overprovi sioning PostgreSQL for peak semantic-search traffic and reduces the engineering effort to tune performance, sharding, indexing, and retry logic.

Options B and C can work, but they require the team to build and maintain embedding pipelines, query embedding generation, vector index management, and operational scaling strategies. Option A does not provide semantic search because it relies on keyword-based matching rather than embeddings.

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.

The least operational overhead approach is to use managed Amazon Bedrock model evaluation workflows with datasets stored in Amazon S3 , and then publish results into Amazon CloudWatch for dashboards. That is exactly what options B and C combine.

Step B correctly places standardized evaluation inputs in Amazon S3 and focuses on granting the evaluation workflow the right permissions to read those datasets. In practice, the key requirement is controlled access to the S3 objects used as evaluation datasets. Establishing IAM permissions and private access patterns (such as using VPC connectivity patterns where applicable to the organization’s networking posture) is aligned with enterprise requirements and avoids building custom storage or data distribution systems for evaluators.

Step C then operationalizes the evaluation lifecycle with minimal infrastructure: a scheduled AWS Lambda function starts evaluation jobs using the S3 dataset location, and a second Lambda function checks job status and pushes results and operational signals to CloudWatch . This meets the platform requirement to surface accuracy metrics in dashboards because CloudWatch metrics/logs can be visualized in dashboards and queried through CloudWatch Logs Insights . It also supports continuous, standardized comparisons across models without requiring developers to run ad-hoc experiments.

The alternatives introduce more operational burden. D and E rely on Amazon SageMaker -based tooling, notebook jobs, and open source evaluation frameworks, which require more environment management, dependency control, scaling considerations, and maintenance over time. A includes CORS, which is primarily a browser-access concern and does not address how Bedrock-managed evaluation jobs securely access S3 in the typical service-to-service pattern.

Therefore, B + C achieves standardized model evaluation, automated scheduling, and dashboard-ready observability with the smallest operations footprint.

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 D is the correct solution because it directly satisfies all three core requirements: data source registration, metadata-based attribution, and end-to-end audit logging, while remaining service-agnostic and scalable across internal and external data sources.

The AWS Glue Data Catalog is the AWS-native service for registering datasets and managing metadata centrally. It supports structured registration of diverse data sources and enables consistent tagging that can be used to attribute generated content back to its original source. This is essential for GenAI applications that combine multiple datasets and must provide traceability for outputs.

Metadata tags applied within the Glue Data Catalog ensure a consistent attribution framework that downstream systems—such as Retrieval Augmented Generation (RAG) pipelines or evaluation systems—can reference without embedding attribution logic directly in application code. This improves maintainability and governance.

AWS CloudTrail provides immutable audit logs of API activity across AWS services, including data access, metadata changes, and pipeline interactions. CloudTrail logs are critical for compliance and regulatory review because they capture who accessed which data, when, and through which service. This satisfies the requirement to maintain audit logs “throughout the pipeline,” not just at storage or application layers.

Option A introduces Lake Formation, which is primarily intended for fine-grained data lake permissions and is not required solely for traceability. Option B relies on CloudWatch Logs, which does not provide authoritative audit logging across services. Option C limits audit scope to S3 access and does not register or govern all data sources comprehensively.

Therefore, Option D provides the most complete and least intrusive solution for traceable, auditable GenAI data pipelines.

Option B is the most appropriate solution because it directly uses Amazon Bedrock cross-Region inference profiles, which are designed to provide resilience and load distribution while respecting data residency boundaries. Cross-Region inference profiles allow applications to distribute inference requests across multiple Regions within a defined geographic boundary, such as Europe or North America, without requiring custom failover logic.

By specifying geographical codes in the inference profile ID, the application ensures that European user data is processed only within Europe-based Regions, satisfying regulatory requirements. At the same time, Bedrock automatically routes requests to healthy Regions within that geography when traffic spikes or service quotas are reached, improving availability and maintaining low latency.

Using separate Amazon API Gateway HTTP APIs for Europe and North America provides a clean, simple routing layer that directs users to the appropriate regional inference profile. This avoids complex custom routing or retry logic in application code and minimizes operational overhead.

Option A relies on custom routing and manual monitoring, which increases complexity and does not provide automatic resilience. Option C introduces custom retry and fallback logic that risks violating data residency requirements if misconfigured. Option D requires significant application-level failover logic and adds operational burden with Global Accelerator configuration.

Therefore, Option B best meets the requirements for low latency, data residency compliance, resilience during traffic spikes, and minimal operational complexity.

Option B is correct because the retrieval problem requires filtering before semantic similarity search, not merely returning metadata after retrieval. Amazon Bedrock Knowledge Bases can apply filters to metadata associated with source documents, so timestamp, author, and research-domain attributes can restrict the candidate set before vector results are returned. This prevents old, wrong-domain, or unauthorized-author documents from entering the retrieved context. AWS also documents S3 knowledge-base metadata files and filterable metadata attributes. Option A only returns metadata with already-selected vector results, so irrelevant documents can still be retrieved. Option C introduces DynamoDB and a two-stage custom retrieval path. Option D adds Comprehend and Aurora, increasing operational complexity. The key design principle is to make the required attributes filterable in the vector retrieval layer. AWS Documentation

================

Option B best meets the requirements because it provides systematic diagnosis, measurable comparison, and historical traceability of prompt performance. By placing prompts under version control and testing them against complex clinical documents, the company can consistently reproduce issues, track regressions, and compare prompt behavior using quantifiable metrics such as factual accuracy, completeness, and consistency. Automated testing ensures scalability and repeatability, while version history preserves prompt evolution over time.

Option A lacks objective metrics and does not address complex documents. Option C focuses on live traffic experimentation but does not inherently diagnose prompt inconsistencies or preserve detailed historical evaluations. Option D adds medical entity analysis but introduces unnecessary service coupling and does not provide robust prompt version history or automated comparative benchmarking. Therefore, Option B is the most complete and disciplined solution.

Option C directly addresses the root cause of truncated and inconsistent responses by using AWS- recommended semantic chunking and dynamic retrieval rather than static or sequential chunk processing. Amazon Bedrock documentation emphasizes that foundation models have fixed context windows and that sending oversized or poorly structured input can lead to truncation, loss of context, and degraded output quality.

Semantic chunking breaks documents based on meaning instead of fixed token counts. By using a breakpoint percentile threshold and sentence buffers, the content remains coherent and semantically complete. This approach reduces the likelihood that important concepts are split across chunks, which is a common cause of inconsistent summarization results.

The RetrieveAndGenerate API is designed specifically to handle large documents that exceed a model’s context window. Instead of forcing all content into a single inference call, the API generates embeddings for chunks and dynamically selects only the most relevant chunks based on similarity to the user query. This ensures that the FM receives only high-value context while staying within its context window limits.

Option A is ineffective because chaining chunks sequentially does not align with how FMs process context and risks exceeding context limits or introducing irrelevant information. Option B improves structure but still relies on larger parent chunks, which can lead to inefficiencies when processing very large documents. Option D processes segments independently, which often causes loss of global context and inconsistent summaries.

Therefore, Option C is the most robust, AWS-aligned solution for resolving truncation and consistency issues when processing large technical documents with Amazon Bedrock.

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.

Option B is the correct solution because it parallelizes independent foundation model inference tasks while maintaining orchestration, observability, and time-bound execution. AWS Generative AI best practices emphasize reducing end-to-end latency by parallelizing independent inference calls rather than scaling individual calls vertically.

In this scenario, each report requires multiple independent analyses such as financial extraction, sentiment analysis, and compliance validation. These tasks do not depend on each other’s output, making them ideal candidates for parallel execution. AWS Step Functions provides a Parallel state that can invoke multiple AWS Lambda functions simultaneously, drastically reducing total processing time compared to sequential execution.

By invoking Amazon Bedrock from separate Lambda functions in parallel, the system can reduce batch execution time from 45 seconds to well under the 10-second requirement, assuming each inference call remains within acceptable latency bounds. Step Functions also provide built-in error handling, retries, and state tracking, which improves reliability without increasing complexity.

CloudWatch metrics allow teams to monitor both workflow execution time and individual model inference latency, enabling performance tuning and operational visibility. Configuring client-side timeouts ensures that slow or failed model invocations do not block the entire batch.

Option A still processes tasks sequentially and therefore cannot meet the strict latency requirement. Option C introduces queuing delays and sequential processing within each report, which increases total execution time. Option D relies on container-based sequential processing and adds unnecessary operational overhead for a workload that is event-driven and latency-sensitive.

Therefore, Option B best meets the performance, scalability, and operational efficiency requirements for high-speed batch document analysis using Amazon Bedrock.

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 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 C is the best answer because it implements a decoupled, queue-based moderation pipeline that invokes the expensive model only when the low-latency model is not confident enough. Amazon SQS is designed to decouple distributed application components and support asynchronous processing. AWS documentation describes SQS as a fully managed message queuing service that enables decoupling and scaling of microservices, distributed systems, and serverless applications. This matches the requirement for high resiliency and sudden demand spikes because incoming requests can be buffered in a durable queue rather than overwhelming the model-processing layer.

Using AWS Fargate to process queue messages is also appropriate because Fargate provides serverless container compute for Amazon ECS or Amazon EKS workloads. It allows the company to run scalable processing workers without managing EC2 capacity directly. AWS Prescriptive Guidance includes architectures that use API Gateway, Amazon SQS, and AWS Fargate to process events asynchronously, which supports the same decoupled processing model required in this question.

The two-stage queue design also optimizes cost. The small, low-latency classifier is used first for all requests. Only requests with confidence below 0.65 are placed into the second queue and processed by the larger model. This avoids running the larger model for every moderation request. High-confidence results can be completed quickly by the first-stage processor, while uncertain results are isolated into an asynchronous second-stage workflow.

Option A is incorrect because it synchronously calls the larger model for low-confidence results, which violates the requirement to process low-confidence requests asynchronously. Option B is incorrect because it invokes both models for every request, increasing cost and eliminating the benefit of confidence-based escalation. Option D requires managing EC2 instances and uses keyword heuristics instead of model confidence, so it is less resilient and less aligned with Bedrock-based moderation. Therefore, option C is correct.

Option C is the correct solution because AWS AppConfig is designed for real-time, validated, centrally managed configuration changes with safe rollout, immediate propagation, and rollback support—exactly matching the company’s requirements.

By storing routing rules, cost thresholds, regulatory constraints, and A/B testing logic in AWS AppConfig, the company can switch between Amazon Bedrock foundation models without redeploying Lambda code. AppConfig supports feature flags, dynamic configuration updates, JSON schema validation, and staged rollouts, which are essential for safely managing complex and frequently changing routing logic.

Using the AWS AppConfig Agent, Lambda functions can retrieve cached configurations efficiently, ensuring low latency even under thousands of concurrent requests. This approach allows the Lambda function to apply proprietary business logic—such as user tier, transaction value, Region compliance, and real-time cost metrics—before selecting the appropriate FM.

Option A is operationally fragile because environment variable changes require function restarts and do not support validation or controlled rollouts. Option B is too limited for complex, dynamic logic and is difficult to maintain at scale. Option D misuses Lambda authorizers, which are intended for authentication and authorization, not high-frequency dynamic routing decisions.

Therefore, Option C provides the most scalable, flexible, and low-overhead architecture for dynamic, regulation-aware FM routing in a global GenAI system.

Option C is the correct solution because AWS AppConfig is purpose-built to manage dynamic application configurations with low latency, strong validation, and minimal operational overhead, which directly matches the company’s requirements.

AWS AppConfig enables the company to centrally manage model selection logic, inference parameters, and customer-tier routing rules without redeploying Lambda functions. By using feature flags, the company can easily perform A/B testing of new models or prompt strategies by gradually rolling out changes to a subset of users or customer tiers. This allows experimentation and controlled releases without code changes.

AppConfig also supports JSON schema validation, which is critical for validating parameters such as temperature, maximum token limits, and other model-specific settings before they are applied. This prevents invalid or unsafe configurations from being deployed and reduces the risk of runtime errors or degraded model behavior in production.

Using the AWS AppConfig Agent allows Lambda functions to retrieve configurations efficiently with built-in caching and polling mechanisms, minimizing latency and avoiding excessive calls to configuration services. This approach scales well for high-throughput, low-latency applications such as GenAI APIs behind Amazon API Gateway.

Option A introduces unnecessary redeployment logic and polling complexity. Option B requires building and maintaining custom configuration access patterns in DynamoDB and does not natively support feature flags or schema validation. Option D adds operational overhead by requiring ElastiCache cluster management and custom validation logic.

Therefore, Option C provides the most scalable, flexible, and low-maintenance solution for dynamic model switching, A/B testing, and safe configuration management in a GenAI application.

Option B best meets the combined resiliency and observability requirements because it applies AWS-recommended retry behavior for transient throttling and enables true distributed tracing across service boundaries. During peak traffic, intermittent failures are commonly caused by throttling and other transient conditions. The AWS SDK standard retry mode provides exponential backoff with jitter, which reduces synchronized retry storms, prevents cascading failures, and improves overall system stability. Jitter is important because it spreads retry attempts over time, reducing load amplification during throttling events.

For observability, AWS X-Ray provides distributed tracing that follows a request across components such as API Gateway or load balancers, application services, and downstream calls to Amazon Bedrock. X-Ray can identify where latency is being introduced and which downstream call is contributing most to end-to-end response time. This is required to “identify latency contributors” and isolate performance issues under load.

The requirement also states that the company must correlate performance issues with specific FM characteristics. X-Ray annotations are designed for this purpose: the application can annotate traces with the model ID, inference parameters, region, or inference profile used. This enables filtering and analysis (for example, comparing latency or error patterns by model, parameter set, or endpoint configuration) without building a separate telemetry system.

Option A’s fixed-delay retries increase synchronized retry behavior and do not provide distributed tracing. Option C does not prevent cascading failures and cannot provide cross-service tracing. Option D is incorrect because CloudTrail is an audit logging service and does not provide distributed tracing for request latency analysis.

Therefore, Option B provides the correct combination of resilient retries and deep, model-correlated distributed observability for Amazon Bedrock workloads.

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 D is the correct solution because it directly addresses all three requirements: high retrieval accuracy, hallucination detection, and reduced human review costs. AWS recommends a layered evaluation strategy for high-stakes domains such as healthcare, where generative outputs must be both accurate and safe.

Using an automated LLM-as-a-judge evaluation enables scalable, consistent assessment of generated responses for factual grounding, relevance, and hallucination risk. This automated screening significantly reduces the number of responses that require manual inspection. Only responses that fall below defined quality thresholds or exhibit ambiguous behavior are escalated to targeted human reviews, which optimizes review effort and cost.

The use of Amazon Bedrock built-in evaluations provides standardized metrics specifically designed for RAG systems, including retrieval precision, faithfulness to source documents, and hallucination rates. These evaluations integrate directly with Amazon Bedrock knowledge bases and models, eliminating the need to build and maintain custom evaluation pipelines.

Option A focuses on entity extraction confidence, which does not reliably detect hallucinations in generative text. Option B requires maintaining and scaling a separate fine-tuned evaluation model, increasing complexity and cost. Option C is useful for regression testing but cannot detect hallucinations in real-world, open-ended clinical queries.

Therefore, Option D provides the most effective and operationally efficient approach to maintaining clinical-grade accuracy while minimizing human review effort.

The combination of Options C and D delivers comprehensive, real-time observability for Amazon Bedrock workloads with the least operational overhead by relying on native integrations and managed services.

Amazon Bedrock publishes built-in CloudWatch metrics for model invocations and token usage. Option C leverages these native metrics directly, allowing teams to build centralized CloudWatch dashboards without additional data pipelines or custom processing. CloudWatch alarms provide threshold-based alerting for token consumption, enabling proactive cost and usage control across all foundation models. This approach aligns with AWS guidance to use native service metrics whenever possible to reduce operational complexity.

Option D complements CloudWatch by enabling advanced, stakeholder-specific visualizations through Amazon Managed Grafana. The zero-ETL integration allows Bedrock and CloudWatch metrics to be visualized directly in Grafana without building ingestion pipelines or managing storage layers. Grafana dashboards are particularly well suited for serving different audiences, such as engineering, finance, and product teams, each with customized views of token usage and trends.

Option A introduces unnecessary complexity by adding a business intelligence layer that is better suited for historical analytics than real-time operational monitoring. Option B is useful for deep log analysis but requires query maintenance and does not provide efficient real-time dashboards at scale. Option E involves multiple services and custom data flows, significantly increasing operational overhead compared to native metric-based observability.

By combining CloudWatch dashboards and alarms with Managed Grafana’s zero-ETL visualization capabilities, the company achieves real-time visibility, flexible dashboards, and automated alerting across all Amazon Bedrock foundation models with minimal operational effort.

This complex set of requirements is best addressed by Amazon Bedrock Prompt Management . It allows the creation of parameterized prompts where variables (like demographics) can be injected at runtime, ensuring consistent medical terminology while adapting recommendations to the specific patient. Prompt Management natively supports versioning and approval workflows , which is a requirement for clinical safety and compliance. For audit and retention, Bedrock model invocation logging can be configured to send detailed prompt and response data to Amazon S3 . Storing these logs in S3 supports the 2-year retention requirement and facilitates local audits. S3 is more cost-effective for long-term storage than CloudWatch Logs alone. CloudTrail (Option A) only logs management events, not the actual prompt/response content required for medical auditing.

Option A is the correct solution because it uses native Amazon Bedrock governance and evaluation capabilities to meet regulatory, performance, and deployment timeline requirements with the least operational overhead.

Amazon Bedrock guardrails provide built-in safety controls that enforce responsible AI policies directly during inference. Custom content filters and toxicity detection protect customer interactions and prevent disallowed investment guidance patterns without requiring custom application logic. Guardrails operate inline and are optimized for low latency, which helps meet the strict 200 ms response-time requirement.

Hallucination detection is addressed through Amazon Bedrock Model Evaluation, which supports automated evaluation at scale using LLM-as-a-judge techniques. This enables the company to detect factual inaccuracies and policy violations systematically, without building custom evaluation pipelines or requiring extensive human review. Evaluation outputs can be surfaced as metrics.

Storing all prompt-response pairs in Amazon DynamoDB provides a low-latency, highly scalable audit store that aligns with financial regulatory requirements. Using TTL enforces data retention policies automatically, reducing compliance risk and storage overhead.

Amazon CloudWatch custom metrics integrate seamlessly with existing compliance dashboards, allowing near–real-time monitoring of safety interventions, hallucination rates, and drift indicators. CloudWatch anomaly detection can be applied to these metrics to surface behavior changes quickly.

Option B relies on custom Lambda logic and S3-based auditing, increasing latency and operational complexity. Option C introduces additional services that increase setup time and may exceed the 60-day deployment window. Option D uses non–Bedrock-native monitoring and adds unnecessary infrastructure layers.

Therefore, Option A provides the most complete, compliant, and low-overhead governance solution for a regulated GenAI financial services application.

Option A best meets the requirements because it directly implements a Retrieval Augmented Generation pattern for conversational recommendations using managed Amazon Bedrock capabilities and a scalable vector store. The company’s source data already resides in Amazon S3, which aligns naturally with Amazon Bedrock Knowledge Bases ingestion workflows. A knowledge base can ingest book text, reviews, and metadata, generate embeddings using a supported embedding model, and persist those vectors in a purpose-built vector backend such as Amazon OpenSearch Service. This enables semantic retrieval that is well suited to conversation-driven intent, where user prompts are often descriptive and do not map cleanly to keyword filters.

The requirement to suggest books based on conversations implies the system must interpret natural language context and retrieve relevant passages, reviews, and metadata to ground the recommendation. Knowledge Bases provide managed orchestration for embedding creation and retrieval, which reduces development effort compared to building custom embedding pipelines. OpenSearch Service provides scalable vector search and k-nearest neighbors style similarity retrieval, which supports low-latency responses when properly indexed and sized.

For scaling to more than 10,000 concurrent users, the API layer design in option A is a common AWS pattern: Amazon API Gateway provides a managed front door with throttling and request handling, while AWS Lambda scales horizontally with demand and can invoke the knowledge base retrieval operations. This separates compute scaling from the vector store scaling and helps keep latency predictable under load.

Option B is not the best choice because DynamoDB is not the standard native vector store target for Amazon Bedrock Knowledge Bases in this context and would introduce additional implementation complexity around vector indexing and similarity search behavior. Option C requires substantial ML lifecycle work, model hosting, tuning, and continuous iteration to achieve quality recommendations at scale. Option D provides strong enterprise search, but it focuses on retrieval and FAQs rather than a managed RAG recommendation workflow grounded in embeddings and conversational context for generative responses.

AWS Step Functions provides native support for human-in-the-loop workflows, making it the best fit for regulatory oversight requirements. The waitForTaskToken integration pattern is explicitly designed to pause a workflow until an external actor—such as a human reviewer—completes a task.

In this architecture, AI-generated recommendations are sent to a human technician for review. The workflow pauses execution using a task token. Once the technician approves or rejects the recommendation, an AWS Lambda function calls SendTaskSuccess or SendTaskFailure, allowing the workflow to continue deterministically.

This approach ensures full auditability, as Step Functions records every state transition, timestamp, and execution path. Storing review outcomes in Amazon DynamoDB provides durable, queryable audit records required for regulatory compliance.

Option A requires custom orchestration and lacks native workflow state management. Option C incorrectly uses AWS Glue, which is not designed for approval workflows. Option D uses caching instead of durable audit storage and introduces unnecessary complexity.

Therefore, Option B is the AWS-recommended, lowest-risk, and most auditable solution for mandatory human review of AI outputs.

Option C cleanly separates rapidly changing routing configuration from application code. AWS AppConfig Agent is the AWS-recommended mechanism for retrieving configuration data and maintains a local cache while asynchronously polling AppConfig for updates. In Lambda environments, the extension periodically refreshes configuration in the background and lets functions retrieve cached values locally with very low latency. The Lambda function can therefore evaluate proprietary rules involving user tier, transaction value, Region, compliance constraints, and current model costs without redeployment. Model identifiers and routing thresholds can change independently through AppConfig while API Gateway exposes one stable application endpoint. Option A requires environment-variable modifications and Lambda configuration updates. Option B makes API Gateway transformation templates responsible for complex business decisions. Option D incorrectly couples model-routing decisions with authorization logic and requires separate model-specific Lambda functions, reducing flexibility. AWS Documentation

================

Options A and C provide the most AWS-native governance approach. Amazon Bedrock Prompt Management supports reusable parameterized prompts and explicit prompt versions, allowing teams to preserve approved configurations and deploy stable versions instead of uncontrolled working drafts. An organizational approval gate can therefore be applied before a new prompt version becomes production-approved. AWS CloudTrail records Amazon Bedrock API activity, including the identity making the request, the operation performed, time, source information, and other event details, providing the required audit trail for Bedrock resource modifications and use. CloudWatch can complement this with operational usage monitoring and reporting. Option B recreates prompt governance through S3, Lambda, and SNS. Option D uses general-purpose configuration storage rather than Bedrock-native prompt management. Option E adds substantial custom state and filtering logic without providing the same managed prompt-version lifecycle. AWS Documentation

================

Amazon Q Developer Pro provides a built-in administrator dashboard designed specifically for organizational observability. This dashboard provides native visibility into user-level metrics across the entire AWS Organization, allowing administrators to identify active vs. underused subscriptions and recognize power users. Crucially, it tracks high-level usage patterns, including code acceptance metrics (such as lines of code generated and accepted), which is a key requirement for measuring ROI and identifying training needs. Using the built-in dashboard provides the necessary insights with the least operational overhead, as it does not require building custom data pipelines (Option C) or complex log processing architectures (Option D).

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.

Amazon CloudWatch provides the most integrated path for unifying technical and business metrics. By importing business metrics into CloudWatch (via custom metrics or metric streams), teams can build custom dashboards that provide a single pane of glass for both system health and conversion performance. Composite alarms allow stakeholders to be notified only when multiple conditions are met (e.g., high latency and dropping conversion rates), reducing alert fatigue. Applying anomaly detection to these metrics is essential for GenAI workloads because performance baselines can shift subtly; CloudWatch can automatically detect these deviations and trigger alerts through Amazon SNS . This solution provides comprehensive correlation and automated alerting with less operational complexity than managing external visualization servers (Option B) or multi-service analytics pipelines (Option C).

Option C is the correct solution because it uses a single, well-tuned Amazon Bedrock guardrail that applies different actions to different content types, which is the recommended approach for minimizing false positives while enforcing strong policy controls.

Setting content filters to medium rather than high reduces overblocking of benign customer conversations while still preventing harmful content. Amazon Bedrock guardrails are designed to balance precision and recall, and medium sensitivity is commonly recommended for customer-facing financial services use cases.

Denied topics explicitly prevent the assistant from discussing investment advice, which is a regulatory requirement. Including definitions and sample phrases improves detection accuracy and reduces ambiguity.

Sensitive information filters support different actions per context. Masking PII in responses preserves conversational usefulness for legitimate customer support while preventing exposure of sensitive data. Blocking sensitive financial information in inputs prevents downstream processing of disallowed content before it reaches the foundation model.

Critically, enabling both input and output evaluation ensures that guardrails are applied consistently at every stage of interaction. Custom blocked messages and audit logging provide clear compliance evidence for regulators and internal audits.

Option A causes excessive false positives by blocking all PII outright. Option B introduces unnecessary complexity and is not how Bedrock guardrails are intended to be applied. Option D uses orchestration logic that Bedrock guardrails already handle natively.

Therefore, Option C best satisfies enforcement, flexibility, auditability, and accuracy requirements.

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 medi cal 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 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 is the optimal solution because it provides scalable semantic search, rich metadata filtering, and tight integration with Amazon Bedrock while minimizing operational overhead. Amazon OpenSearch Serverless is designed for high-volume, low-latency search workloads and removes the need to manage clusters, capacity planning, or scaling policies.

With support for vector search and structured metadata filtering, OpenSearch Serverless enables efficient similarity search across 10 million embeddings while applying constraints such as language, publication date, regulatory agency, and document type. This is critical for financial services use cases where relevance and compliance depend on precise filtering.

Integrating OpenSearch Serverless with Amazon Bedrock Knowledge Bases enables a fully managed RAG workflow. The knowledge base handles embedding generation, retrieval, and context assembly, while Amazon Bedrock generates responses using a foundation model. This significantly reduces custom glue code and operational complexity.

Multilingual support is handled at the embedding and retrieval layer, allowing documents in English, Spanish, and Portuguese to be searched semantically without language-specific query logic. OpenSearch’s distributed architecture ensures consistent low-latency responses for real-time customer interactions.

Option B increases operational overhead by requiring database tuning and scaling for vector workloads. Option C does not support advanced metadata filtering, which is a key requirement. Option D introduces unnecessary complexity and is not optimized for large-scale semantic document retrieval.

Therefore, Option A best meets the requirements for performance, scalability, multilingual support, and minimal management effort in an Amazon Bedrock–based RAG application.

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.

AWS Glue Data Catalog and its associated crawlers are the standard AWS tools for automatic discovery and centralized cataloging of datasets. For the generated summaries, storing them in Amazon S3 allows the use of object tags for metadata (like source IDs), making them easily queryable. The critical requirement for " tamper-evident, immutable audit logs " is met by enabling Bedrock model invocation logging to an S3 bucket protected by S3 Object Lock (compliance mode). To further guarantee that logs have not been altered, AWS CloudTrail log file integrity validation uses cryptographic hashes to provide non-repudiation and a verifiable audit trail. This combination covers data management, metadata attribution, and high-standard security compliance.

Option B best meets all requirements because it combines AWS-recommended resiliency patterns for transient failures with streaming-aware handling and adaptive protection against cascading retries during peak load. When timeouts and throttling occur, naïve retries can amplify traffic and worsen outages. Exponential backoff with jitter is the standard AWS best practice because it spreads retry attempts over time, reduces synchronized retry storms, and lowers the probability of repeatedly colliding with service limits.

The requirement also states the strategy must “adapt to changing service availability” and “prevent overwhelming Amazon Bedrock.” A circuit breaker pattern directly addresses this by temporarily stopping or reducing retries when failure rates exceed a threshold, allowing the system to degrade gracefully instead of continually hammering the service. This is a key mechanism to prevent cascading failures during throttling events.

Because the application uses response streaming and experiences broken chunks, the retry strategy must be streaming-aware. A streaming response handler that detects chunk delivery timeouts and buffers already received chunks prevents the user from losing progress when a connection drops. Resuming from the last successfully received chunk minimizes redundant generation and reduces additional load on the model compared with restarting the entire stream. This supports better user experience and better service efficiency during intermittent failures.

Token-aware request handling is supported in this architecture because the application can apply token budgeting before invoking the model (for example, trimming or summarizing excessive context) while still preserving streaming output behavior. Option B provides the correct backbone for this by focusing on adaptive control and robust streaming recovery.

Option A is too simplistic and risks retry storms. Option C combines conflicting elements (global token limit, cached completions for streaming) and includes impractical “request only missing chunks” behavior that is not a reliable property of streamed generative output. Option D includes useful ideas (load shedding) but relies on static caps and does not provide as strong adaptive retry control as circuit breaking.

Therefore, Option B is the most correct and operationally safe strategy for peak-load Bedrock streaming workloads.

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.

Amazon SageMaker Serverless Inference is specifically designed for applications that experience intermittent or bursty traffic. It automatically scales compute capacity based on the number of requests and scales down to zero when there is no traffic, satisfying the requirement to optimize resource usage during low demand. To meet the 500 ms latency requirement during peak periods and avoid " cold start " delays, provisioned concurrency keeps a specified number of execution environments warm and ready to respond immediately. This provides a balance between the cost-effectiveness of serverless and the performance predictability of provisioned instances. Multi-model endpoints (Option A) can introduce " noisy neighbor " issues and latency spikes, while asynchronous inference (Option D) is intended for long-running workloads and cannot meet sub-500 ms 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 correct because the application is currently invoking the base foundation model identifier, which routes traffic to the on-demand capacity pool rather than the company’s purchased provisioned throughput. In Amazon Bedrock, provisioned throughput is attached to a specific provisioned resource created through the provisioned throughput APIs. To consume that reserved capacity, inference requests must target the provisioned resource identifier that represents the purchased throughput, not the generic model identifier used for on-demand inference.

The code snippet uses modelId= " anthropic.claude-v2 " . This value selects the on-demand endpoint for that model. As a result, requests are subject to on-demand quotas and throttling behavior, while the provisioned throughput remains idle. This directly explains the CloudWatch observation: provisioned capacity metrics show unused capacity because no traffic is being directed to the provisioned resource, and the on-demand path is throttling because it is exceeding the applicable on-demand limits during peak volume.

Replacing the modelId value with the provisioned throughput ARN returned by the CreateProvisionedModelThroughput workflow ensures the runtime invocation is routed to the reserved capacity. Once traffic is directed correctly, the purchased model units provide the consistent throughput required for predictable performance during business hours, which is exactly why provisioned throughput is used.

Option A could increase capacity, but it does not fix the core issue that the application is not using the provisioned resource at all. Option C can reduce the impact of throttling temporarily, but it adds latency and does not guarantee consistent throughput; it also still wastes the provisioned capacity. Option D changes the response delivery mechanism, but throttling is a capacity routing and quota issue, not a streaming API issue.

Option B matches the defining requirement: the company, rather than a managed agent harness, must own the orchestration loop . Amazon Bedrock AgentCore Runtime provides a serverless, purpose-built environment for hosting agents and tools while remaining framework agnostic. AWS explicitly supports agents created with frameworks such as LangGraph, Strands, and CrewAI as well as fully custom agents that use no agent framework.

AWS further distinguishes AgentCore Runtime from its managed harness by stating that with Runtime, the customer brings the agent code and “the orchestration loop is yours.” The custom application can therefore implement proprietary tool-selection logic, complex conversation-state transitions, specialized retry algorithms, proprietary telemetry, validation stages, and any other control-flow behavior required by the business. Runtime supplies infrastructure capabilities such as isolation, scaling, sessions, authentication, and observability plumbing rather than taking ownership of the decision loop.

Amazon Bedrock can separately provide managed foundation-model inference, allowing the company to use Bedrock FMs without deploying or operating model-serving infrastructure.

A is incorrect specifically because of its final statement. AgentCore Policy is an authorization mechanism, not an orchestration or retry engine. AWS documents that Policy evaluates Cedar or Dogwood policies to determine whether tool invocations are permitted or forbidden. It can govern which actions are authorized, but it does not implement an agent ' s tool-selection algorithm or failure-retry strategy.

C relinquishes precisely the orchestration control the company requires by depending on managed behavior. D could provide deterministic workflow coordination but places Step Functions outside the agent loop and is cumbersome for highly dynamic multi-turn agent reasoning, proprietary state transitions, and tool-selection decisions.

Therefore, B preserves complete application-level orchestration control while outsourcing scalable runtime infrastructure and foundation-model inference to AWS.

==========

This requirement explicitly calls for character-by-character streaming, long-running responses, low latency, and massive concurrency, which aligns directly with Amazon Bedrock streaming inference patterns.

Amazon Bedrock provides the InvokeModelWithResponseStream API specifically for streaming partial model outputs as tokens are generated. This enables near-instant feedback to users instead of waiting for the full response to complete, which is essential when responses last up to 45 seconds.

Amazon API Gateway WebSocket APIs are purpose-built for bidirectional, low-latency, server-initiated communication, allowing the backend to push characters or tokens to clients in real time. This eliminates inefficient polling and supports thousands of concurrent open connections.

AWS Lambda integrates natively with WebSocket APIs and scales automatically with connection volume, enabling a fully managed, serverless architecture. This approach maintains security, centralized authentication, throttling, and observability while avoiding direct client access to Bedrock APIs.

Option B introduces polling latency and unnecessary API overhead and does not provide true streaming. Option C violates AWS security best practices by exposing Bedrock directly to clients and does not scale securely. Option D only serves completed responses and cannot meet the real-time streaming requirement.

Therefore, Option A is the only solution that fully satisfies streaming behavior, concurrency, latency, and managed-service constraints.

Option B is the correct solution because it delivers end-to-end governance, security, and auditability across Amazon Bedrock, Amazon Rekognition, and the underlying data layer while meeting strict privacy and compliance requirements.

Using IAM attribute-based access control (ABAC) allows the company to control access to foundation models and data based on department, role, or workload attributes rather than static permissions. This is critical for controlling FM access patterns at scale. Enforcing specific ModelId and GuardrailIdentifier values with IAM condition keys ensures that only approved models and guardrails are used, which directly supports secure model I/O and governance requirements.

Configuring VPC endpoints for Amazon Bedrock ensures that all model invocations remain on private AWS network paths, reducing data exfiltration risk and supporting privacy standards. AWS CloudTrail captures both management and data events, providing a definitive audit trail of who accessed which resources and when. Sending logs to CloudTrail Lake enables centralized, long-term, queryable auditing across services.

Amazon S3 server access logging adds file-level visibility into video archive access, which is essential for compliance and forensic analysis. Amazon CloudWatch alarms provide near real-time detection of anomalous or unauthorized activity across Amazon Bedrock, Amazon Rekognition, and AWS KMS.

Option A focuses primarily on model-level tracing but lacks comprehensive IAM governance and S3 access auditing. Option C provides partial controls but lacks identity-aware auditing and model governance. Option D focuses on anomaly detection and classification but does not explicitly control FM access patterns.

Therefore, Option B best satisfies all stated requirements in a unified, auditable, and security-first architecture.

Option B is the best fit because hierarchical chunking is designed to preserve local detail while keeping broader document context available during retrieval, which directly addresses the problem of questions spanning methodology, results, and discussion. In large scientific papers, a single answer often depends on linked paragraphs across adjacent sections. If the knowledge base retrieves only small, isolated chunks, the RAG system can cite text that is semantically close to a query term but not contextually correct, producing inconsistent answers and irrelevant citations.

With hierarchical chunking, the knowledge base creates child chunks that are small enough for high-precision vector similarity matching, such as 200 tokens, which improves the likelihood that the retrieved text is tightly related to the user’s query. At the same time, each child chunk is associated with a larger parent chunk , such as 1,000 tokens, which retains the surrounding narrative and section-level context. This structure helps the retrieval pipeline return passages that include the relevant subsection plus the explanatory framing that prevents misinterpretation, which is especially important in scientific writing where methods, results, and discussion are interdependent.

The configured overlap further reduces boundary effects where key statements split across chunks. This improves continuity for paragraphs that bridge sections, such as a results paragraph that references the methodological setup or a discussion paragraph interpreting a specific metric.

Option A can improve consistency slightly, but fixed-size chunking still risks separating related paragraphs and does not provide a built-in mechanism to retrieve broader context linked to precise matches. Option C can create more meaningful boundaries, but it does not guarantee the parent-level context that hierarchical chunking provides at retrieval time. Option D increases operational burden and is not practical at the scale of 25 million