Free AWS Certified Generative AI Developer - Professional (AIP-C01) Sample Questions

Correct answer: C — Implement a retry strategy with exponential backoff and jitter to gradually increase the wait time between retries and randomize the delay.

• A. Incorrect. An immediate retry does not allow the service any time to recover from the high load that caused the throttling. This approach can create a 'thundering herd' effect, where multiple clients bombard the service simultaneously, exacerbating the problem and leading to more failures.
• B. Incorrect. While better than an immediate retry, a fixed-interval strategy can lead to synchronized retries. Multiple clients that receive a throttling exception at the same time might retry in lockstep, causing repeated spikes in traffic that continue to overwhelm the service and prolong the recovery period.
• C. Correct. This is the industry best practice and recommended approach by AWS for handling throttling. Exponential backoff systematically increases the wait time between successive retries, giving the service a progressively longer window to recover. Adding jitter (a small, random amount of time) to the backoff delay prevents clients from retrying in synchronized waves. This combination effectively smooths out request bursts and maximizes the chance of success.
• D. Incorrect. This is an operational response, not an application-level retry strategy. Programmatically increasing Provisioned Throughput on every exception is not a scalable, timely, or cost-effective solution. The scaling process is not instantaneous, and this reactive approach could lead to significant over-provisioning and increased costs. The correct approach is to handle transient errors in the application with a proper retry strategy and use monitoring to make informed, strategic decisions about capacity adjustments.

Correct answer: A — Implement a Circuit Breaker pattern using a Lambda function or DynamoDB state to track failures. If the failure threshold is reached, fail immediately or return a fallback response without calling the API.

• A. This is the correct approach. The Circuit Breaker pattern is specifically designed for this scenario. It tracks the number of consecutive failures from a service. Once a threshold is reached, the 'circuit opens,' and subsequent calls fail immediately (fail-fast) or return a fallback response without attempting to contact the failing service. This prevents the Step Function from stalling, reduces costs from repeated failed API calls, and protects the downstream service. After a configured timeout, the circuit moves to a 'half-open' state to test the service before closing the circuit and resuming normal operations.
• B. This is incorrect and would worsen the problem. Aggressively retrying with a high number of attempts and a short interval would hammer the failing third-party service, significantly increasing API call costs and prolonging the Step Function execution time during the outage. This is the opposite of the requirement to stop sending requests after failures.
• C. This is incorrect. An SQS Dead-Letter Queue (DLQ) is used to store messages that could not be processed successfully after a number of retries. It's a pattern for ensuring eventual consistency and debugging failed events, but it does not prevent the Step Function task from repeatedly attempting to call the failing service in real-time. The initial calls would still occur and fail, which does not meet the requirement to stop sending requests immediately.
• D. This is incorrect. AWS Shield Advanced is a managed Distributed Denial of Service (DDoS) protection service. Its purpose is to safeguard applications against malicious traffic floods, not to handle application-level failures or outages from a third-party dependency. It provides no functionality for implementing a circuit breaker pattern.

Correct answer: C — Temperature

• A. Incorrect. The Learning Rate is a hyperparameter used during model training, not inference. It controls the step size for updating the model's weights during optimization with algorithms like gradient descent and does not affect the output's randomness at inference time.
• B. Incorrect. Batch Size refers to the number of samples processed in a single iteration. It is primarily a hyperparameter for model training that can impact training speed and performance. While batching can be used during inference for throughput, it does not control the creativity or randomness of the generated output.
• C. Correct. Temperature is a key inference parameter available in Amazon Bedrock that directly controls the randomness of the model's output. A higher temperature value (e.g., >1.0) increases randomness, making the output more diverse and creative. A lower temperature value (e.g., <1.0, approaching 0) makes the model's output more deterministic and focused on the most probable tokens.
• D. Incorrect. Epochs define the number of times an entire training dataset is passed through the learning algorithm during the training phase. This is a fundamental concept in model training and is not an adjustable parameter during inference to control output characteristics.

Correct answer: C — Configure Amazon Bedrock Guardrails with a denied topics filter for 'cryptocurrency speculation' and a sensitive information filter for PII. Attach this guardrail to the model invocation.

• A. Incorrect. While system prompts and few-shot examples can guide a model's behavior, they are not a reliable enforcement mechanism. The model might still fail to adhere to the instructions, and this method does not guarantee that the check happens before the model processes the input, which is a key compliance requirement.
• B. Incorrect. This approach introduces significant architectural complexity and latency. Invoking an AWS Lambda function for preprocessing and another service like Amazon Comprehend for post-processing adds multiple network hops. Furthermore, using regular expressions for PII detection can be brittle and incomplete, and this custom solution is less robust and auditable than a native, integrated feature.
• C. Correct. Amazon Bedrock Guardrails are the purpose-built feature for this exact use case. They provide a declarative way to enforce policies with minimal latency. A 'denied topics' filter can block queries about 'cryptocurrency speculation' before they are sent to the model. A 'sensitive information' filter can detect and redact PII in both the user's input and the model's generated response, ensuring compliance at both stages of the interaction.
• D. Incorrect. This is an overly complex solution for a policy enforcement requirement. Amazon Bedrock Agents and orchestration flows are designed for multi-step task completion, not for applying universal safety and compliance filters. This method would introduce unnecessary latency, complexity, and potential points of failure compared to the streamlined and purpose-built Guardrails feature.

Correct answer: B — Amazon Bedrock Prompt Flows.

• A. Incorrect. While AWS Lambda and Amazon EventBridge (or AWS Step Functions) can be used to build serverless orchestration workflows, they are general-purpose services. This approach is not optimized for visually designing, interactively testing, and debugging generative AI prompt chains. It would require significant custom code to manage the state, logic, and calls to the Bedrock API, lacking the specialized, Bedrock-native tooling for prompt engineering.
• B. Correct. This option describes the concept of a tool specifically designed for building complex generative AI applications on Bedrock. Features like Agents for Amazon Bedrock allow developers to create multi-step tasks, orchestrate calls to models, and implement business logic. This approach provides a managed, and often visual, way to compose, test, and deploy workflows with prompt chaining and conditional logic, abstracting away the underlying orchestration complexity.
• C. Incorrect. Amazon SageMaker Model Monitor is a tool for monitoring deployed machine learning models. Its purpose is to detect data quality issues and concept drift post-deployment to ensure model performance over time. It is not involved in the design, development, or execution of prompt-based workflows.
• D. Incorrect. AWS Glue is a serverless data integration and ETL (Extract, Transform, Load) service. It is designed for large-scale data processing and preparation, not for orchestrating real-time, interactive generative AI workflows that involve prompt chaining and conditional logic based on model outputs.

Correct answer: A — Implement a Hybrid Search architecture using OpenSearch that combines k-NN vector search scores with Okapi BM25 keyword scores, applying normalization and weights to boost exact keyword matches.

• A. This is the correct approach. A Hybrid Search architecture combines the strengths of both semantic (vector/k-NN) search and lexical (keyword/BM25) search. This allows the system to find semantically similar items for natural language queries while using the BM25 score to strongly rank and prioritize exact matches on specific fields like Product IDs or brand names. By normalizing the scores from both search types and applying custom weights, a developer can ensure that an exact match on a keyword field is always ranked at the top, directly solving the stated problem.
• B. This is incorrect. While increasing embedding dimensionality can sometimes improve the nuance of semantic representation, it is not a reliable method for enforcing exact matches on specific identifiers like SKUs. Embedding models are designed to capture meaning and context, not to memorize arbitrary strings. The problem is better solved by using a search mechanism designed for exact matches, like keyword search, rather than attempting to force this behavior from embeddings.
• C. This is incorrect. Switching the vector store from OpenSearch to Aurora with pgvector, or changing the approximate nearest neighbor algorithm (e.g., from HNSW to IVFFlat), only addresses the vector search component of the system. This does not solve the core problem, which is the need to combine semantic search with a mechanism for prioritizing exact keyword matches. The issue is not with the vector search's accuracy but its inability to handle a non-semantic, exact-match requirement on its own.
• D. This is an inefficient and overly complex solution. Using a Bedrock Agent to decompose the query adds an extra LLM call, which increases both latency and cost. Furthermore, after executing two separate searches (one for keywords, one for vectors), the developer would still need to implement a custom logic to merge and rank the results, which is the original problem. A hybrid search feature within OpenSearch accomplishes this in a single, optimized query.

Correct answer: B — Input: A query and a list of documents/chunks; Output: The same documents ordered by relevance score.

• A. This describes the function of a text generation or summarization model, not a reranker. A reranker's purpose is to reorder an existing list of documents based on relevance, not to generate new text like a summary.
• B. This is the correct description of a reranker's function. In a RAG workflow, after the initial retrieval from a vector store, the reranker takes the original user query and the retrieved documents as input. It then re-evaluates each document's semantic relevance to the query and outputs the same set of documents, but reordered from most to least relevant, along with relevance scores. This helps prioritize the most pertinent information for the final generation step.
• C. This describes a vector similarity calculation, which is a core part of the initial retrieval step in a RAG system, often performed by a vector database. The database calculates the distance (e.g., Euclidean distance or cosine similarity) between the query vector and document vectors to find the nearest neighbors. A reranker operates on the retrieved text content, not directly on the vectors, in a subsequent step.
• D. This describes a technique known as query expansion. Query expansion involves augmenting the original query with synonyms or related terms to broaden the search and improve the initial retrieval step. This is a separate process from reranking, which evaluates and reorders the documents *after* they have been retrieved.

Correct answer: C — Use the Amazon OpenSearch Ingestion (OSI) pipeline to automatically replicate data from DynamoDB to an OpenSearch Service index for vector search.

• A. Incorrect. While this approach is functional, it requires the developer to build, manage, and scale a custom streaming pipeline using DynamoDB Streams and AWS Lambda. This increases operational overhead compared to a fully managed ingestion solution, which contradicts the requirement to minimize management.
• B. Incorrect. This solution involves a complex, multi-step batch process (Export, Process, Load). Additionally, while Amazon Neptune has vector search capabilities, it is primarily a graph database and is not the most optimized or cost-effective choice for a high-read, pure vector similarity search use case compared to a specialized service like OpenSearch.
• C. Correct. This is the most appropriate solution as it directly addresses all requirements. The Amazon OpenSearch Ingestion (OSI) pipeline is a fully managed service that automates data synchronization from DynamoDB to OpenSearch Service, minimizing management overhead. Amazon OpenSearch Service is designed for high-read search workloads and has robust, scalable k-NN vector search capabilities, making it ideal for the "find similar products" functionality.
• D. Incorrect. This is an anti-pattern for performance at scale. DynamoDB lacks native vector search indexes. Performing a table scan for every similarity search request would be extremely slow and costly, failing to meet the 'extremely high read traffic' requirement.

Correct answer: C — Amazon OpenSearch Serverless (Vector Engine).

• A. Incorrect. The provisioned version of Amazon OpenSearch Service is not a serverless solution. It requires you to manage clusters, instances, and shard topology. While Auto Scaling can help manage capacity, it does not eliminate the operational overhead of managing the underlying infrastructure, which contradicts the core requirement for a serverless architecture.
• B. Incorrect. Although Amazon Aurora has an architecture that separates storage and compute, it is not a fully serverless offering as you still need to provision and manage database instances and read replicas. Furthermore, while the `pgvector` extension provides vector capabilities, Aurora is a general-purpose relational database, not a specialized vector engine optimized for high-performance Approximate Nearest Neighbor (ANN) search at the scale of billions of vectors.
• C. Correct. Amazon OpenSearch Serverless is a fully managed, serverless offering with a purpose-built Vector Engine. It is designed to scale to billions of vectors and completely abstracts away the management of instances, clusters, and shards. Its architecture decouples storage and compute, allowing indexing and search capacity to scale independently and automatically, which perfectly aligns with all the requirements.
• D. Incorrect. Amazon DynamoDB is a serverless NoSQL database, but it lacks native vector search capabilities. It is not designed for similarity or Approximate Nearest Neighbor (ANN) searches. A Global Secondary Index (GSI) is used to speed up queries on non-key attributes, not for performing vector-based nearest neighbor searches.

Correct answer: B — Chain two Amazon Bedrock calls. Use a lightweight text model (e.g., Amazon Titan Text Express) with a system prompt to rewrite and expand the user's input into a descriptive prompt. Pass the output to the Titan Image Generator.

• A. This is incorrect. While a frontend spell-checker can fix typos, it fails to address the core problem of ambiguous or underspecified descriptions. It cannot enrich or expand the prompt with the necessary detail required for high-quality image generation.
• B. This is the correct approach. Chaining two Amazon Bedrock calls is an efficient pattern that meets all requirements. A lightweight, fast text model can be guided by a system prompt to correct spelling, resolve ambiguity, and creatively expand the user's simple input into a rich, descriptive prompt. This enhances the input for the Titan Image Generator, leading to better images while minimizing latency and development overhead by staying within the managed Bedrock environment.
• C. This is incorrect. Amazon Comprehend is designed for natural language understanding and analysis, not generative rewriting. Extracting key phrases would likely strip the prompt of essential context and nuance, resulting in a terse, fragmented input that would degrade, rather than improve, the quality of the generated image.
• D. This is incorrect as it directly violates the constraints. Training, deploying, and maintaining a custom model on Amazon SageMaker introduces significant development overhead, cost, and operational complexity. The latency would also likely be higher than a serverless Bedrock API call. Using a pre-trained, readily available text model in Bedrock is a far more efficient and practical solution.

Correct answer: A — Use RAG (Retrieval Augmented Generation). Index the table schemas and descriptions in a vector store. When a user asks a question, retrieve only the schemas of the top-k most relevant tables and insert them into the prompt.

• A. Correct. Retrieval Augmented Generation (RAG) is ideal for providing large external knowledge to a model. By indexing table schemas and descriptions in a vector store, only the top-k most relevant schemas are retrieved based on the user's question and inserted into the prompt. This keeps the prompt within context limits, gives the model focused context, and scales easily as the schema changes.
• B. Incorrect. Removing vowels from column names would only marginally reduce token count, not enough to fit a massive schema within context limits. More critically, it corrupts the semantic meaning of table and column names, making it nearly impossible for the model to understand the schema and generate accurate SQL queries.
• C. Incorrect. Fine-tuning is not a reliable method for memorizing large, factual datasets like a database schema; it is computationally expensive and time-consuming. Additionally, any schema change would require a costly re-fine-tuning, whereas a RAG-based approach allows quick and inexpensive updates to the vector index.
• D. Incorrect. Splitting the request into 200 separate API calls prevents the model from understanding relationships between tables, making joins impossible. This would produce fragmented, incorrect query snippets, and the orchestration overhead would introduce significant latency without yielding a coherent SQL query.

Correct answer: B — Purchase and configure Amazon Bedrock Provisioned Throughput for the specific model units required.

• A. Incorrect. The On-Demand tier for Amazon Bedrock uses shared capacity and does not provide guarantees against throttling, especially during a massive, predictable traffic surge. While implementing exponential backoff with retries is a best practice for handling transient errors, it is a reactive measure that mitigates throttling after it occurs, rather than proactively preventing it. This would lead to a degraded user experience for a mission-critical launch.
• B. Correct. Amazon Bedrock Provisioned Throughput is the ideal solution for this scenario. It is specifically designed for applications with predictable, high-throughput requirements. By purchasing and configuring a specific number of Model Units (MUs), you reserve dedicated inference capacity, which guarantees a consistent level of throughput and prevents throttling exceptions. This directly addresses the core requirement for guaranteed capacity for a high-stakes launch.
• C. Incorrect. Amazon SageMaker Serverless Inference is designed for workloads with intermittent or unpredictable traffic patterns and is not suitable for sustained, high-throughput needs. It does not provide the hard guarantees of reserved capacity that Provisioned Throughput does and is subject to its own concurrency limits and potential cold starts. For a predictable, massive launch, this option would be unreliable.
• D. Incorrect. A cross-region architecture primarily addresses high availability, disaster recovery, and latency reduction for a global user base. While it distributes traffic, it does not inherently guarantee capacity or prevent throttling within each region. Each region would still be subject to its own capacity limits. To handle the required load, you would still need to use Provisioned Throughput in each region, making Provisioned Throughput the fundamental solution, not the cross-region deployment itself.

Correct answer: B — Agents for Amazon Bedrock

• A. Incorrect. AWS Glue Studio is a visual interface for creating, running, and monitoring extract, transform, and load (ETL) jobs. It is designed for data integration and transformation pipelines, not for orchestrating multi-step generative AI workflows that involve calling language models and executing dynamically generated code like SQL.
• B. Correct. Agents for Amazon Bedrock provide a managed experience to build generative AI applications that can perform multi-step tasks. It allows developers to visually create a flow (often called an orchestration or prompt flow) that chains together calls to foundation models, executes API calls (like running a query against a database), and uses the results to generate a final response. This aligns perfectly with the requirement for a visual interface to build a 'Talk to Data' feature with minimal boilerplate code.
• C. Incorrect. Amazon SageMaker Pipelines is an MLOps tool for orchestrating machine learning workflows, such as data preprocessing, model training, tuning, and deployment. While it is a powerful orchestration tool, it is not designed as a visual prompt engineering or agent-building interface and would require extensive custom development to implement the described interactive flow.
• D. Incorrect. AWS App Runner is a fully managed service for deploying and running containerized web applications and APIs. It is a compute and deployment service, not a development tool for visually building the application logic. While App Runner could host the final application, it does not provide the visual interface to build and test the generative AI flow itself.

Correct answer: B — Use TLS 1.2+ for transit encryption. Configure the Bedrock model customization job or Knowledge Base to use a Customer Managed Key (CMK) from AWS KMS.

• A. Incorrect. All AWS service API calls, including those to Bedrock, are encrypted in transit using TLS by default. AWS Certificate Manager (ACM) is not directly configured by the end-user for the Bedrock service endpoint. Most importantly, this option suggests using AWS-managed keys for storage, which directly contradicts the requirement for the firm to manage the encryption keys themselves.
• B. Correct. This option correctly addresses both requirements. All communication with the Amazon Bedrock API endpoint is secured using TLS 1.2+ for encryption in transit. For encryption at rest with customer-controlled keys, services like Bedrock model customization jobs and Knowledge Bases can be configured to use a Customer Managed Key (CMK) from AWS Key Management Service (KMS). This gives the legal firm direct control over the key's lifecycle and access policies, fulfilling all stated requirements.
• C. Incorrect. This approach is fundamentally flawed. If the application encrypts the payload before sending it to Bedrock, the generative AI model will receive unintelligible ciphertext instead of the plaintext contract data. The model cannot process encrypted data, rendering the service useless for its intended purpose. This method does not address how data is encrypted at rest within the Bedrock service either.
• D. Incorrect. A VPC Endpoint for Bedrock provides a private network connection, ensuring traffic does not traverse the public internet, but TLS is still the mechanism that provides encryption in transit over that connection. This option fails to address the primary requirement of using a customer-managed key for data stored at rest by Bedrock for processes like fine-tuning or within a Knowledge Base. Only encrypting the final output in an S3 bucket is insufficient.

Correct answer: B — Use an API Gateway WebSocket API. Configure the integration to trigger a Lambda function that utilizes the `InvokeModelWithResponseStream` API and writes chunks back to the connection ID.

• A. This approach fails both requirements. The `InvokeModel` API is synchronous, waiting for the entire response before returning, which prevents the token-by-token 'typewriter' effect. Additionally, a standard REST API is request-response and does not provide the persistent, bi-directional channel needed for the user to interrupt the generation mid-stream.
• B. This is the correct solution. An API Gateway WebSocket API establishes a persistent, bi-directional communication channel between the client and the backend. This allows the Lambda function to stream response chunks from the `InvokeModelWithResponseStream` API back to the client for the 'typewriter' effect, while also allowing the client to send an interrupt message back to the server over the same connection to stop the generation.
• C. While this approach correctly uses `InvokeModelWithResponseStream` to enable streaming, Server-Sent Events (SSE) is a uni-directional protocol (server-to-client). It cannot satisfy the requirement for bi-directional communication, as the client cannot send an interrupt message back to the server over the same SSE connection. A separate communication channel would be required, making it less suitable than WebSockets.
• D. This architecture is unsuitable and overly complex. It uses the non-streaming `InvokeModel` API, failing the primary requirement. Furthermore, using DynamoDB and AppSync subscriptions introduces significant latency (write -> trigger -> publish) which is not appropriate for real-time, token-by-token streaming. It also does not offer a direct bi-directional communication channel for interruptions.

Correct answer: B — Use Amazon API Gateway as a proxy. Configure a VTL (Velocity Template Language) mapping template in the Integration Request to transform the incoming XML payload into the required JSON format before sending it to the backend Lambda.

• A. This approach is incorrect because introducing an Amazon SQS queue adds an asynchronous buffering step. This increases complexity, end-to-end latency, and cost due to the additional SQS and Lambda invocations. This pattern is suitable for decoupling systems or handling bursts, not for minimizing latency in a synchronous-style request.
• B. This is the correct and most efficient solution. Amazon API Gateway's VTL mapping templates can perform payload transformations directly at the gateway layer. This converts the incoming XML to the required JSON format before forwarding the request to the backend, without needing a separate compute service for the transformation. This method minimizes both compute cost and latency, directly addressing the core requirements of the problem.
• C. While this approach is functionally viable, it is not optimal. Using a Lambda function to parse the XML introduces additional compute costs for the transformation logic and is subject to potential cold-start latency. Compared to performing the transformation natively in API Gateway, this option is less efficient in terms of both cost and speed.
• D. This solution is incorrect as it is overly complex and expensive for the task. AWS Step Functions and AWS Glue are powerful services for orchestrating complex workflows and large-scale ETL jobs, respectively. Using them for a simple, per-request payload transformation is excessive and would result in significantly higher latency and cost.

Correct answer: B — Use Amazon Bedrock Prompt Flows. Use the 'Prompt' node for LLM calls and the 'Lambda' node for custom logic, relying on the managed environment of Prompt Flows.

• A. Incorrect. This option requires building and managing a custom container image, which the team explicitly wants to avoid. Additionally, AWS Lambda does not provide a visual builder for orchestrating prompt chains.
• B. Correct. Amazon Bedrock Prompt Flows provides a visual, node-based interface for building complex prompt chains. It includes a managed environment with nodes for LLM calls and custom Python logic, eliminating the need to manage Lambda layers or container images.
• C. Incorrect. Amazon ECS Fargate requires packaging the application into a container image, violating the constraint against managing container images. It also lacks a visual builder and introduces more infrastructure management overhead than a managed orchestration service.
• D. Incorrect. Amazon SageMaker Notebook instances are designed for interactive development, not for running managed, production-grade workflows. This solution requires manual instance management and does not offer a visual builder for prompt chaining.

Correct answer: B — Use a model cascading (chaining) approach: Route all queries to a smaller, cheaper model first, and only invoke the larger model if the confidence score is below a threshold.

• A. Incorrect. Deploying the high-intelligence model on SageMaker with Auto Scaling adjusts capacity based on traffic, but it does not differentiate between simple and complex queries. Every request is routed to the expensive model, failing to optimize costs for the 80% of simple FAQs.
• B. Correct. A model cascading approach routes all queries to a smaller, cheaper model first, and only invokes the larger model when the confidence score is below a threshold. This ensures the high-cost model is used only for the 20% of complex queries, directly addressing the cost-optimization requirement.
• C. Incorrect. Provisioned Throughput in Amazon Bedrock provides a lower hourly rate for consistent workloads, but it does not enable selective routing based on query complexity. All queries would still be processed by the high-intelligence model, including the simple FAQs, negating the cost savings.
• D. Incorrect. Deploying the high-intelligence model on AWS Lambda and routing all queries to it does not solve the routing problem; every request still uses the expensive model. Additionally, large models often exceed Lambda's resource limits, and cold starts can impact performance, making this unsuitable for the scenario.

Correct answer: B — Implement a Step Functions 'Catch' block for parsing errors that transitions to a 'Repair' task. This task prompts the model with the error message and the previous invalid output to request a correction.

• A. Incorrect. This describes a manual remediation process, not a self-healing mechanism. A self-healing system should recover automatically without human intervention. Relying on an operator introduces significant latency and defeats the purpose of automation.
• B. Correct. This is the most effective and robust solution. AWS Step Functions has built-in error handling using 'Catch' blocks. By catching a specific error (e.g., a JSON parsing exception), the workflow can transition to a corrective state. This 'Repair' task can then re-invoke the model with a modified prompt that includes the original invalid output and the error message, asking it to fix the format. This creates an automated, in-workflow self-correction loop.
• C. Incorrect. Simply retrying the 'InvokeModel' task is unlikely to succeed. Retries are effective for transient issues like network timeouts, not for deterministic problems like malformed output. Without changing the prompt or providing corrective feedback, the model will likely produce the same invalid JSON on subsequent attempts.
• D. Incorrect. First, no Foundation Model can realistically guarantee 100% JSON compliance in all situations. Second, this is an architectural change, not a dynamic, in-workflow self-healing mechanism. A resilient workflow should be designed to handle potential failures at runtime, regardless of the specific model being used.

Correct answer: B — It provides a standardized interface for connecting Foundation Models to data sources and tools, decoupling the tool implementation from the model provider.

• A. Incorrect. The Model Context Protocol (MCP) standardizes how models and tools exchange messages and context, but it does not perform automatic fine-tuning of foundation models. Fine-tuning involves updating model weights through a separate training process, which is not a function of MCP.
• B. Correct. MCP's primary technical benefit is providing a standardized interface for connecting foundation models to data sources and tools, which decouples tool implementation from the model provider. This promotes interoperability, allowing developers to swap models or tools with minimal integration changes, making the system more modular and maintainable.
• C. Incorrect. MCP does not increase the effective context window of Amazon Bedrock models or compress prompt data through a preprocessing layer. The context window size is an architectural property of the model, and while techniques like RAG can manage context, MCP itself does not perform token reduction or window expansion.
• D. Incorrect. MCP is a protocol for tool interaction and content exchange, not an authentication mechanism. It does not eliminate the need for IAM authentication or establish a direct trust relationship that bypasses AWS credential management; secure access controls remain essential.

Correct answer: B — Implement Amazon Bedrock Agent traces to capture the Chain of Thought (CoT), action inputs, action outputs, and the final response, then store these traces for audit.

• A. Incorrect. While enabling Amazon CloudWatch Logs is useful for general monitoring, it does not provide the structured, step-by-step visualization of the model's reasoning required for a compliance audit. Metric filters are designed to extract numerical metrics from logs, not to reconstruct the agent's complex decision-making process.
• B. Correct. Amazon Bedrock Agent traces are specifically designed for this purpose. They provide a detailed, structured record of the agent's execution, including the model's reasoning (Chain of Thought or CoT), which action groups and tools were invoked, the inputs provided to them, their outputs, and the final response. Storing these traces provides a comprehensive audit trail that meets transparency and explainability requirements for risk officers.
• C. Incorrect. Amazon Macie is a data security service used to discover and protect sensitive data, such as personally identifiable information (PII), in Amazon S3. It does not provide any insight into the operational logic or internal reasoning of an AI model, making it unsuitable for auditing the agent's orchestration process.
• D. Incorrect. This approach relies on the model to self-report its own logic, which is unreliable for compliance and auditing purposes. The model-generated summary could be incomplete, inconsistent, or an inaccurate representation of the actual steps taken. A robust audit solution requires a system-level trace of events rather than a model's own interpretation of its actions.

Correct answer: B — The model's AI Service Card (Model Card).

• A. Incorrect. The AWS Service Health Dashboard provides information about the operational status, availability, and outages of AWS services. It does not contain model-specific details regarding safety testing, intended use cases, or limitations.
• B. Correct. An AI Service Card, often referred to as a Model Card, is a standardized document designed to provide transparency. It details a model's intended use cases, performance characteristics, limitations, and the results of safety testing, including assessments for harmful or sensitive content like self-harm. This is the primary resource for evaluating a model's suitability for responsible AI applications.
• C. Incorrect. Amazon CloudWatch provides operational and performance metrics, such as latency, invocation counts, error rates, and resource utilization. It is used for monitoring the model in production but does not provide any qualitative information about its safety, ethical considerations, or intended use cases.
• D. Incorrect. AWS Artifact is a service that provides access to AWS's security and compliance reports (e.g., SOC, PCI, ISO). These reports pertain to the compliance of the underlying AWS infrastructure and services, not the behavioral characteristics or safety evaluations of a specific Foundation Model.

Correct answer: B — Define a comprehensive OpenAPI schema for the Action Group, including types, descriptions, and required fields, to guide the agent's reasoning and parameter extraction.

• A. This approach is reactive, not preventative. Relying solely on the model's reasoning is insufficient for strict compliance, and a `try-catch` block in the Lambda function only handles errors after an invalid call has already been made. It does not prevent the agent from generating incorrect parameters or data types in the first place.
• B. This is the correct and recommended approach for Bedrock Agents. An OpenAPI schema provides a formal, machine-readable contract for the Action Group. By precisely defining parameters, data types, required fields, and descriptions, it gives the agent a strict structure to follow. This enables the Bedrock Agent service to correctly extract and validate parameters before invoking the Lambda function, directly preventing hallucinations and enforcing schema compliance at the source.
• C. This method is significantly less reliable than a formal schema. While providing instructions in a text prompt can offer some guidance, it relies on the model's interpretation of natural language rather than programmatic enforcement. This approach lacks a guaranteed validation step and is still highly susceptible to hallucinations, malformed outputs, and incorrect data types.
• D. This is an overly complex, inefficient, and reactive solution. Implementing an external validation mechanism using EventBridge adds latency and architectural complexity. It only intercepts and validates the API call after the agent has already generated it. The native functionality of using an OpenAPI schema within the Action Group is the most direct, integrated, and effective method for proactive validation.

Correct answer: B — Implement Amazon Comprehend's real-time analysis API with PII detection, configured to redact detected PHI entities, then pass the output to the LLM.

• A. Incorrect. Amazon Macie is designed for discovering and protecting sensitive data at rest, primarily in Amazon S3, and is not suitable for real-time analysis of streaming text. This approach would require custom integration to buffer the text, trigger Macie, and then use a Lambda function for masking, which introduces significant latency and custom code maintenance, directly contradicting the requirements.
• B. Correct. Amazon Comprehend provides a real-time `DetectPiiEntities` API that can identify and redact Personally Identifiable Information (PII), which includes Protected Health Information (PHI). By using this managed service's API, the developer can send the chat text, receive a redacted version in real-time, and then pass the safe text to the SageMaker LLM. This approach directly addresses the real-time redaction requirement and minimizes custom code maintenance as it leverages a purpose-built AWS service.
• C. Incorrect. AWS Glue and its FindMatches ML transform are designed for large-scale batch data processing, data integration, and entity resolution (like finding duplicate records). They are not suitable for low-latency, real-time analysis of a chat stream and would fail to meet the core requirements.
• D. Incorrect. Amazon Bedrock Guardrails are a feature of the Amazon Bedrock service and can only be applied to models invoked through the Bedrock API. They cannot be directly applied to a custom LLM hosted on a separate Amazon SageMaker endpoint, making this an architecturally infeasible solution for the given scenario.

Correct answer: B — Enable Model Invocation Logging in Amazon Bedrock settings, pointing to an S3 bucket with Object Lock enabled. AWS Bedrock does not use customer inference data to train base models by default.

• A. This is incorrect. Amazon Bedrock does not automatically log invocation payloads to CloudWatch Logs; Model Invocation Logging must be explicitly configured. Furthermore, there is no generic 'Do Not Track' setting in the AWS support center that controls data usage for Bedrock. Data privacy is managed through specific service configurations and account-level opt-out policies.
• B. This is the correct solution. Enabling Model Invocation Logging in Amazon Bedrock is the designated feature for creating an audit trail of prompts and responses. Directing these logs to a customer-controlled S3 bucket meets the centralization requirement. Applying S3 Object Lock on the bucket ensures the logs are immutable and retained for the required 7-year period. This also aligns with the privacy requirement, as AWS Bedrock's default policy is not to use customer inference data for training its base models.
• C. This is incorrect. VPC Flow Logs capture metadata about IP traffic, such as source/destination IP addresses, ports, and byte counts. They do not capture the application-layer payload, which contains the actual text of the prompts and model outputs. Therefore, this method cannot be used for auditing the content of the chatbot interactions.
• D. This is incorrect. While deploying a model on Amazon SageMaker and using its Data Capture feature is a valid way to log payloads for that service, the question explicitly states the solution must use Amazon Bedrock. Switching to a different service is not the correct approach and is unnecessary as Bedrock has a native feature to meet the requirements.

Correct answer: B — AWS CloudTrail Management Events

• A. Incorrect. Amazon OpenSearch Service Audit Logs are designed to track user activity and events *within* the OpenSearch Service itself, such as user authentication, document access, and search queries. They do not provide an audit trail for AWS management API calls that modify the service's configuration, nor do they capture the IAM principal that initiated such a change.
• B. Correct. AWS CloudTrail is the definitive AWS service for governance, compliance, and auditing of account activity. CloudTrail Management Events record all management API calls made to AWS services, including Amazon OpenSearch Service. These logs contain the essential information for an audit, such as the identity of the API caller (the IAM principal), the timestamp of the call, the source IP address, and the specific configuration change made, directly answering who changed the settings and when.
• C. Incorrect. Amazon CloudWatch Vended Logs are service-generated logs (e.g., application logs, slow query logs, error logs) from services like OpenSearch that are published to CloudWatch for operational monitoring and troubleshooting. They are not the primary mechanism for auditing management-level configuration changes or identifying the IAM principal responsible.
• D. Incorrect. AWS Config tracks the configuration state of AWS resources over time and provides a history of *what* changed and when the resource's state was modified. However, it does not directly identify *who* made the API call to initiate the change. For that level of detail, AWS Config often needs to be used in conjunction with CloudTrail, making CloudTrail the primary source for identifying the actor.

Correct answer: A — It provides a centralized view of all models, endpoints, and monitoring jobs, allowing governance officers to audit model behavior, drift status, and lineage info in one place.

• A. Correct. SageMaker Model Dashboard provides a centralized view of all models, endpoints, and monitoring jobs, enabling governance officers to audit model behavior, drift status, and lineage information in one place. This unified visibility is essential for maintaining oversight and compliance within an AI governance framework.
• B. Incorrect. SageMaker Model Dashboard is an observability tool, not an automated retraining system. While it can detect performance degradation or drift, triggering a retraining pipeline and replacing production endpoints must be implemented separately using services like SageMaker Pipelines or AWS Step Functions.
• C. Incorrect. This describes Amazon SageMaker Ground Truth, which uses human annotators to label raw training data. SageMaker Model Dashboard does not perform data labeling; it focuses on monitoring and auditing deployed models.
• D. Incorrect. SageMaker Model Dashboard does not generate deployment code or manifests for Kubernetes. Model deployment, including container definitions and deployment manifests, is handled through the SageMaker SDK, AWS CLI, or other infrastructure-as-code tools.

Correct answer: A — Use Amazon Rekognition Content Moderation API to analyze the image. If the moderation labels for 'Violence' or 'Explicit' exceed the threshold, block the request before invoking Bedrock.

• A. Correct. Amazon Rekognition Content Moderation API is purpose-built to detect inappropriate content such as violence and explicit material in images. By analyzing the image and blocking requests when moderation labels for 'Violence' or 'Explicit' exceed a threshold, it creates a specialized, auditable pre-check layer before invoking Bedrock.
• B. Incorrect. Amazon Comprehend is an NLP service for analyzing text, and base64 encoding of an image does not contain semantic information about visual content. Using entity recognition and sentiment analysis on the encoded string is ineffective for detecting violent or explicit imagery.
• C. Incorrect. Bedrock Guardrails' text filters are designed for text-based inputs and outputs, not for directly moderating visual content in multi-modal prompts. The requirement for a specialized, auditable pre-check layer calls for an explicit step before invoking Bedrock, not reliance on internal model processing or text-only guardrails.
• D. Incorrect. Using Amazon Mechanical Turk for real-time human verification of every image introduces significant latency, is not scalable, and can be costly. It also raises privacy and compliance concerns by exposing user data to external reviewers, making it unsuitable for a chatbot's pre-check layer.

Correct answer: B — Use Amazon Bedrock Batch Inference (or SageMaker Batch Transform) to process the requests in bulk, avoiding the overhead of always-on infrastructure.

• A. This option is incorrect. Amazon SageMaker Real-time Inference is designed for low-latency, online predictions and requires persistent, always-on infrastructure. This model is expensive and unnecessary for a one-time batch workload where latency is not a concern. Sending requests sequentially would also be extremely slow and inefficient for 500,000 items.
• B. This is the correct approach. Services like Amazon Bedrock Batch Inference or Amazon SageMaker Batch Transform are specifically designed for large-scale, asynchronous, offline inference workloads. They provision compute resources only for the duration of the job and then automatically shut them down. This model avoids the cost of idle, always-on endpoints, efficiently processes data in parallel, and is the most cost-effective solution for this use case.
• C. This option is suboptimal. While using Bedrock On-Demand with a custom script is functionally possible, it's less efficient and more costly for a large batch job compared to a dedicated batch service. This approach involves making 500,000 individual API calls, which can lead to higher per-inference costs, potential API throttling, and requires custom logic to manage concurrency, retries, and error handling.
• D. This option is incorrect because it introduces significant operational overhead and cost risks. Manually provisioning a powerful and expensive EC2 instance like a P4d, deploying the model, and managing the process is complex. Most importantly, relying on manual termination is risky; forgetting to shut down the instance would lead to substantial, unnecessary costs. Managed services like SageMaker Batch Transform are more cost-effective and operationally simpler for this scenario.

Correct answer: B — Subscribe to Amazon EventBridge events for `Bedrock Knowledge Base Ingestion Job Status Change` and alert on failures or long durations.

• A. This is incorrect. Monitoring the S3 `PutObject` metric only confirms that a file has been uploaded to the S3 bucket. It provides no visibility into the subsequent Bedrock Knowledge Base ingestion process, such as its status, duration, or whether it succeeded or failed. This metric cannot help diagnose delays happening after the initial upload.
• B. This is the correct approach. Amazon Bedrock Knowledge Base is integrated with Amazon EventBridge and emits events for ingestion job status changes (e.g., `STARTING`, `IN_PROGRESS`, `COMPLETE`, `FAILED`). By subscribing to these events, the operations team can create rules to trigger notifications or alerts on failures or when a job exceeds a predefined duration threshold. This provides direct, real-time, and actionable insights into the health and performance of the ingestion pipeline.
• C. This is an incorrect and inefficient method. Polling the vector database frequently to count vectors puts unnecessary load on the database, is not scalable, and is an indirect way of monitoring. It doesn't provide specific diagnostic information about why an ingestion job is slow or has failed, and it may not accurately reflect the status of in-progress or partially failed jobs.
• D. This is incorrect. S3 Server Access Logs are primarily for security and access auditing, not real-time operational monitoring. These logs are delivered with a significant delay, are costly to parse at scale, and only indicate that the Bedrock service principal read an object. They do not provide structured information about the ingestion job's status, duration, or error details.

Correct answer: B — Reduce the `ef_search` (examination factor) and `M` (max connections) parameters in the HNSW index settings.

• A. Incorrect. An HNSW index is an Approximate Nearest Neighbor (ANN) index designed for high-speed queries on large datasets by trading perfect accuracy for speed. A flat index performs a brute-force, exhaustive search across all vectors. Switching to a flat index would dramatically increase query latency for 10 million vectors, which is the opposite of the desired outcome.
• B. Correct. Tuning HNSW index parameters is the standard method for managing the trade-off between query speed and recall. Reducing `ef_search` (a query-time parameter for search depth) and `M` (a build-time parameter for graph connectivity) decreases the thoroughness of the search. This directly reduces query latency at the cost of lower recall, which perfectly aligns with the stated business priority.
• C. Incorrect. Increasing the dimensionality of embedding vectors generally worsens performance. It increases the index size, memory consumption, and the computational cost of calculating distances (a phenomenon known as the 'curse of dimensionality'). This would lead to higher query latency, not lower.
• D. Incorrect. Pre-filtering applies metadata filters before the vector search, reducing the search space. Post-filtering performs the vector search first and then filters the results. While the better choice depends on the specific workload and filter selectivity, switching to post-filtering is not a guaranteed optimization and can often be slower. More importantly, it is not the primary and most direct method for trading recall for speed, unlike tuning HNSW parameters.

Correct answer: A — Cause: Index fragmentation. Solution: Trigger a force merge or index refresh operation on the vector store.

• A. This is the correct answer. In many vector databases, frequent updates and deletes don't immediately remove data but mark it for deletion, leading to index fragmentation and staleness. This degrades the quality of the index over time, causing search accuracy metrics like recall to drop, even while query speed (latency) remains acceptable. Operations like force merge, compaction, or re-indexing are standard maintenance procedures to consolidate index segments, permanently remove deleted documents, and rebuild the index structure, which restores retrieval accuracy and recall.
• B. This is incorrect. The scenario explicitly states that latency remains low. High query latency is not the problem being observed. Adding read replicas is a solution for scaling read throughput and reducing latency under high load, but it would not address the underlying issue of a degraded index causing poor recall.
• C. This is incorrect. While embedding drift (concept drift) can cause recall to drop over a long period, it is a less likely cause than index fragmentation given the clue of 'frequent updates and deletes.' Index fragmentation is a direct, mechanical consequence of data churn. Retraining an embedding model is a significant, resource-intensive operation and is usually performed on a longer cadence to address fundamental shifts in data distribution, not as a routine fix for index health.
• D. This is incorrect. Context window overflow is an issue related to the Large Language Model (LLM) that processes the retrieved documents, not the vector database itself. Furthermore, increasing the `top_k` parameter simply retrieves more documents. While this might coincidentally find a relevant document that was previously missed, it's a workaround that doesn't fix the root cause of the degraded index and may negatively impact precision and increase costs.

Correct answer: B — Data Leakage (Contamination); mitigate by performing n-gram overlap analysis between training and evaluation sets and removing overlaps.

• A. Incorrect. While the symptoms are similar to overfitting (good performance on known data, poor on unknown), the root cause is more specific. Overfitting is about a model learning the training data too closely, including its noise. The issue here is data leakage. Furthermore, increasing the inference temperature only makes the model's output more random; it does not address the fundamental problem of a compromised evaluation set.
• B. Correct. This scenario is a classic example of data leakage, also known as data contamination. It occurs when data from the evaluation set is present in the training data, leading to inflated and misleadingly optimistic performance metrics. The correct mitigation is to decontaminate the datasets. This involves auditing the training and evaluation sets to find and remove overlaps. Techniques like n-gram overlap analysis, fuzzy matching, and data fingerprinting are used to identify and eliminate these contaminating samples, ensuring the evaluation set remains a true holdout set.
• C. Incorrect. Catastrophic forgetting describes a model losing previously learned knowledge when it is fine-tuned on a new task or dataset. This is not what is happening in the scenario. Fine-tuning the model on the evaluation set would be counterproductive, as it would explicitly train the model on the test data, completely invalidating the evaluation process.
• D. Incorrect. Concept drift occurs when the statistical properties of the data distribution change over time, causing a model's performance to degrade in production. While this can cause poor production performance, the specific cause identified in the question is data leakage, not changing data patterns. Retraining on newer articles is a strategy to combat concept drift but does not fix the underlying issue of evaluation data contamination.

Correct answer: B — Switch to a keyword-based (lexical) search or a hybrid search strategy that weights keyword matches for acronyms higher than semantic matches.

• A. This is incorrect. Increasing the chunk overlap size is a technique used to maintain semantic context between adjacent document chunks. It helps the model understand information that spans chunk boundaries but does not solve the core problem of the embedding model's inability to understand out-of-vocabulary terms like specific internal acronyms.
• B. This is the correct solution. The root cause is that a purely semantic (vector) search fails when queries contain specific terms, like internal acronyms, that are not well-represented in the pre-trained embedding model's vocabulary. A keyword-based (lexical) search excels at finding exact string matches. A hybrid search strategy combines the strengths of both: it uses lexical search to accurately retrieve documents containing the specific acronyms while still using semantic search for the conceptual parts of the query, providing the most robust and relevant results.
• C. This is incorrect. Lowering the similarity threshold will return more documents, but it does not address the fundamental issue that the query embedding for the acronym is not meaningful. This approach would likely increase the number of irrelevant documents (noise) and reduce overall precision, making it a poor and unreliable workaround.
• D. This is incorrect. This approach fundamentally undermines the purpose of a RAG system, which is to augment an LLM with external knowledge. The LLM's internal knowledge base was not trained on the company's private data and therefore will not contain information about internal acronyms. Relying on it would almost certainly lead to factual inaccuracies or hallucinations.

Correct answer: B — The percentage of relevant documents found in the top K retrieved results.

• A. This is incorrect. Recall@K is a measure of retrieval effectiveness, focusing on the coverage of relevant items. It does not measure the speed or latency of the query. Metrics like response time or query throughput are used to evaluate the speed of a system.
• B. This is the correct definition. Recall@K measures the fraction of all truly relevant documents that are successfully retrieved within the top K results for a given query. It is a key metric for evaluating how well a search system finds relevant items, helping to diagnose cases where relevant information is missed by the retrieval system.
• C. This is incorrect. Recall@K is a standard information retrieval performance metric focused on relevance. The cost per query is a separate financial or operational metric used for budgeting and resource management, and has no direct relationship with the concept of recall.
• D. This is incorrect. The presence of duplicate vectors is an issue related to data quality and index maintenance. While duplicates might indirectly affect search results, Recall@K is not designed to measure them. Its specific purpose is to quantify how many ground-truth relevant documents are returned in the top K results.