Option B is the correct solution because Amazon Bedrock provisioned throughput is only used when the application explicitly invokes the provisioned model ARN, not the base foundation model ID. In the provided code, the application is calling the standard model identifier (anthropic.claude-v2), which routes requests to on-demand capacity instead of the purchased provisioned throughput.

When the CreateProvisionedModelThroughput API is used, Amazon Bedrock returns a provisioned model ARN that represents the reserved capacity. Applications must reference this ARN in the modelId parameter when invoking the model. If the base model ID is used instead, Bedrock treats the request as on-demand traffic, which explains why CloudWatch metrics show unused provisioned capacity alongside throttled on-demand requests.

Option A would increase capacity but would not fix the root cause because the application is not using the provisioned resource at all. Option C adds resiliency but does not ensure usage of provisioned throughput and would still incur throttling. Option D changes the response delivery mechanism but does not affect capacity routing.

Therefore, Option B directly resolves the throttling issue by correctly routing traffic to the reserved capacity and ensures that the company benefits from the provisioned throughput it has purchased.

Option C best satisfies the requirement for rapid, correlated detection of model-related performance degradation. Amazon CloudWatch Application Insights provides automated observability across application components running on Amazon EC2, identifying abnormal behavior patterns without requiring extensive manual configuration.

Using custom metrics for recommendation quality, token usage, and response latency allows the company to directly monitor FM behavior, not just infrastructure health. Applying dimensions such as request type and user segment enables fine-grained correlation between performance issues and specific customer interactions or workloads.

CloudWatch anomaly detection is critical because it establishes dynamic baselines from historical data and detects deviations automatically. This enables alerts to be generated within minutes when FM behavior changes unexpectedly, satisfying the 10-minute alerting requirement without static thresholds that can miss subtle degradations.

CloudWatch Logs Insights complements metrics by enabling rapid analysis of log patterns, error messages, or unusual request flows associated with degraded recommendations. Because all data remains within CloudWatch, correlation between metrics, logs, and alerts is straightforward and operationally efficient.

Option A focuses on infrastructure metrics and lacks behavioral baselining. Option B provides tracing but not automated anomaly detection. Option D adds significant operational overhead and ingestion complexity for a use case already well supported by CloudWatch-native features.

Therefore, Option C delivers the most effective, scalable, and low-overhead observability solution for detecting FM-related performance deviations.

Option C best addresses both core problems: hallucinated recommendations that do not exist in the catalog and slow response times, while keeping operational overhead low. The most direct way to prevent the model from recommending unavailable products is to ground generation on authoritative product catalog data at inference time. An Amazon Bedrock knowledge base is designed for this pattern by ingesting domain data, chunking content, creating embeddings, and retrieving the most relevant catalog entries when a user asks for recommendations. Implementing Retrieval Augmented Generation ensures the foundation model receives only approved, catalog-backed context and can cite or base its output on those retrieved items. This sharply reduces the likelihood of inventing products, because the response is conditioned on retrieved catalog records rather than relying on the model’s parametric memory.

The requirement also notes that most interactions are unique. That makes response caching far less effective, because there are fewer repeated prompts to benefit from cached outputs. Instead, improving the retrieval and model invocation path is the better optimization. Using the PerformanceConfigLatency parameter set to optimized prioritizes lower latency behavior for model inference, helping meet faster recommendation generation without requiring the company to build and operate additional infrastructure.

The other options do not solve the root cause as reliably. Prompt engineering and streaming can improve perceived latency, but they do not guarantee catalog-only recommendations because the model can still hallucinate items. Guardrails can help detect or block certain undesired outputs, but without consistent catalog grounding they do not ensure every recommendation is derived from the company’s product data. Building a custom OpenSearch validation and caching layer increases operational complexity, and caching is misaligned with predominantly unique interactions.

Alright, after comparing List B (txt file) against List A (Word file) , I have identified the unique questions. These questions cover scenarios or architectural configurations that were not present in the existing list.

Here are the unique questions from List B, formatted as requested:

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 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

Option B provides the least operational overhead because it keeps the solution primarily inside managed Amazon Bedrock capabilities, minimizing custom orchestration code and infrastructure to operate. The core requirements are domain grounding, reduced semantic dilution for complex questions, and consistent low-latency responses at high request volume. A Bedrock knowledge base is purpose-built for Retrieval Augmented Generation by ingesting domain documents, chunking content, generating embeddings, and retrieving the most relevant passages at runtime. This directly addresses the need to reference domain-specific medical terminology from authoritative documents to reduce ambiguity and improve factual accuracy.

Reducing semantic dilution typically requires improving the retrieval query so that the retriever focuses on the most relevant concepts, especially for long or multi-intent questions. Enabling query decomposition allows the system to break a complex medical query into smaller, more targeted sub-queries. This increases retrieval precision and recall for each sub-question, which helps the model generate a more accurate synthesized response grounded in the retrieved medical context.

Amazon Bedrock Flows provide a managed way to orchestrate multi-step generative AI workflows, such as preprocessing the input, performing retrieval against the knowledge base, invoking a foundation model, and formatting the final response. Because flows are managed, the company avoids maintaining custom state machines, multiple Lambda functions, or bespoke routing logic. This reduces operational overhead while still supporting repeatable, observable execution.

Compared with the alternatives, option A introduces an agent plus API Gateway routing and multiple model choices, increasing configuration and runtime complexity. Option C requires hosting and scaling custom models on SageMaker AI, which adds significant operational burden and latency risk. Option D relies on multiple Lambda functions orchestrated by an agent, which adds more moving parts and increases cold-start and integration overhead. Option B most directly meets the requirements with the smallest operational footprint.

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.