AWS Certified AI Practitioner AIF-C01 AIF-C01 Questions 451–525 | Page 7/12

451

MCQmedium

A marketing agency wants to analyze customer feedback from social media posts to gauge sentiment. They have no labeled data and limited ML expertise. The team needs a managed service that provides pre-trained models for sentiment analysis without requiring them to train or manage infrastructure. They also need to process text in multiple languages. Which AWS service should they use?

A.Use Amazon Comprehend with its default sentiment analysis model

B.Use Amazon SageMaker to train a custom sentiment analysis model

C.Use AWS Glue to build a custom NLP pipeline

D.Use Amazon Rekognition for text analysis

AnswerA

Amazon Comprehend provides pre-trained sentiment analysis via a managed API, requiring no labelled data, training, or infrastructure management. Its built-in multilingual support handles social media text across languages, matching the agency's lack of ML expertise and need for immediate sentiment detection.

Why this answer

Amazon Comprehend is a fully managed natural language processing (NLP) service that provides pre-trained models for sentiment analysis, key phrase extraction, and language detection. It requires no labeled data, no model training, and no infrastructure management, making it ideal for teams with limited ML expertise. Comprehend natively supports multiple languages, including Spanish, French, German, and many others, directly addressing the requirement to process text in multiple languages.

Exam trap

The trap here is that candidates may confuse Amazon Rekognition (image/video analysis) with text analysis services, or assume that any AWS ML service (like SageMaker or Glue) can handle NLP tasks without recognizing the specific managed service designed for unstructured text.

How to eliminate wrong answers

Option B is wrong because Amazon SageMaker is a platform for building, training, and deploying custom ML models, which requires labeled data, ML expertise, and infrastructure management—contradicting the requirements for a pre-trained, managed service with no training. Option C is wrong because AWS Glue is a serverless data integration and ETL service, not an NLP service; it cannot perform sentiment analysis or provide pre-trained models. Option D is wrong because Amazon Rekognition is a computer vision service for analyzing images and videos, not for text analysis or sentiment detection.

Full explanation →

452

MCQeasy

A retail analytics team wants an assistant that answers questions about last quarter's sales using data stored in an Amazon S3 bucket of PDF reports and CSV exports. They want the model to cite the underlying documents and avoid inventing figures. Which Amazon Bedrock capability should they use?

A.Guardrails for Amazon Bedrock configured with a denied topics policy.

B.Model evaluation jobs that score responses against a ground-truth dataset.

C.Provisioned Throughput purchased for the foundation model.

D.Knowledge Bases for Amazon Bedrock with the S3 bucket as a data source.

AnswerD

Knowledge Bases for Amazon Bedrock ingests documents from sources such as Amazon S3, chunks and embeds them into a vector store, and retrieves relevant passages at query time so the model grounds answers in those documents and can return citations. This directly satisfies the requirement to answer from the sales reports and avoid fabricated figures.

Why this answer

Grounding answers in private documents requires retrieval, not filtering or capacity. Knowledge Bases for Amazon Bedrock manages ingestion, chunking, embedding, and retrieval from the S3 source, so responses are generated from the sales reports and can include citations. Evaluation, Guardrails, and Provisioned Throughput address quality measurement, content safety, and capacity respectively, none of which supply the source data.

Exam trap

The trap here is confusing content filtering or capacity reservation with retrieval-augmented grounding of private data.

Full explanation →

453

MCQmedium

A company uses Amazon Bedrock to generate content and wants to prevent the model from producing harmful or biased responses. Which AWS service should they configure to enforce content safety policies?

A.Amazon Comprehend for toxicity detection

B.Amazon SageMaker Clarify

C.AWS WAF to filter model responses

D.Amazon Bedrock Guardrails

AnswerD

Guardrails applies configurable content filters and denied-topic policies directly to Bedrock model invocations, blocking harmful or biased output before it reaches the application. This satisfies the requirement to enforce content safety policies at inference time, which prompt engineering alone cannot guarantee.

Why this answer

Amazon Bedrock Guardrails provides content filtering, PII redaction, topic restrictions, and other safety controls. It is purpose-built for content safety in Bedrock.

Full explanation →

454

MCQeasy

A developer is creating a generative AI application using Amazon Bedrock and needs to ensure that responses do not include toxic or harmful content. Which feature should be enabled?

A.Amazon CloudWatch Logs for prompt logging.

B.Amazon Virtual Private Cloud (VPC) for network isolation.

C.Amazon Bedrock Guardrails.

D.AWS Identity and Access Management (IAM) policies.

AnswerC

Guardrails applies configurable content filters and denied-topic policies to both prompts and responses, blocking toxic or harmful output at inference time. This satisfies the requirement that responses never include harmful content, without retraining or prompt engineering.

Why this answer

Amazon Bedrock Guardrails is the correct feature because it is specifically designed to enforce content policies, filter toxic or harmful content, and block undesirable topics in generative AI responses. It provides configurable thresholds for hate, insults, sexual content, violence, and other harmful categories, ensuring compliance with safety requirements without modifying the underlying model.

Exam trap

The trap here is that candidates often confuse monitoring/logging services (CloudWatch) or security controls (VPC, IAM) with content safety features, not realizing that Bedrock Guardrails is the only option that directly filters toxic or harmful content at the application layer.

How to eliminate wrong answers

Option A is wrong because Amazon CloudWatch Logs for prompt logging captures and stores logs for monitoring and debugging, but it does not actively filter or block toxic content in responses. Option B is wrong because Amazon Virtual Private Cloud (VPC) provides network isolation and security at the infrastructure layer, but it has no mechanism to inspect or control the semantic content of AI-generated responses. Option D is wrong because AWS Identity and Access Management (IAM) policies control authentication and authorization for API calls, but they cannot enforce content safety rules or filter harmful language in model outputs.

Full explanation →

455

MCQmedium

A retail company wants to forecast product demand at the SKU level for the next 12 weeks. Which AWS service is purpose-built for this task?

A.Amazon Personalize

B.Amazon Forecast

C.Amazon SageMaker

D.Amazon Kendra

AnswerB

Amazon Forecast is purpose-built for time-series forecasting, automatically handling seasonality, trends and related datasets to produce SKU-level demand predictions. It removes the need to build and tune custom models, directly meeting the 12-week SKU-level forecasting requirement.

Why this answer

Amazon Forecast is a fully managed service purpose-built for time-series forecasting, using machine learning to analyze historical data and predict future demand. It is specifically designed for tasks like SKU-level product demand forecasting over a defined time horizon, such as 12 weeks, without requiring deep ML expertise.

Exam trap

The trap here is that candidates may confuse Amazon Personalize (which also uses ML for predictions) with forecasting, but Personalize is for recommendation systems, not time-series demand prediction.

How to eliminate wrong answers

Option A is wrong because Amazon Personalize is a recommendation engine for personalizing user experiences (e.g., product recommendations, content ranking), not for time-series forecasting of demand. Option C is wrong because Amazon SageMaker is a general-purpose ML platform for building, training, and deploying custom models, requiring manual setup of forecasting algorithms and infrastructure, whereas Amazon Forecast is a managed service optimized for forecasting. Option D is wrong because Amazon Kendra is an intelligent search service for enterprise document retrieval and question answering, not designed for numerical time-series prediction.

Full explanation →

456

Multi-Selecthard

A company uses a text generation model to produce legal documents. They want to minimize the environmental impact of training and inference. Which THREE approaches should they consider?

Select 3 answers

A.Train the model on a smaller but representative dataset

B.Store all training data in multiple AWS Regions for redundancy

C.Use a more efficient model architecture like a distilled version of a larger model

D.Perform all training on on-demand instances to ensure consistent performance

E.Use Amazon SageMaker with managed spot training to reduce idle compute

AnswersA, C, E

A smaller representative dataset shortens each training epoch and reduces total compute hours, lowering the energy consumed during training. This directly satisfies the stem's environmental-impact goal, provided the sample preserves the legal-domain patterns the model must learn.

Why this answer

Option A is correct because training on a smaller but representative dataset reduces the total compute cycles, energy consumption, and carbon emissions required for model training while still maintaining acceptable model quality. Option C is correct because distilled or otherwise more efficient architectures require fewer parameters and FLOPs per inference, directly lowering the energy and hardware resources needed for both training and serving. Option E is correct because SageMaker managed spot training uses spare AWS capacity and automatically checkpoints/resumes jobs, which improves utilization and reduces idle compute, thereby lowering the environmental footprint.

Option B is not appropriate because replicating training data across multiple AWS Regions increases storage, replication traffic, and energy use without reducing the model's environmental impact. Option D is not appropriate because on-demand instances do not inherently reduce energy consumption and can leave resources idle; they address performance consistency, not environmental efficiency.

Exam trap

AIF-C01 often tests the misconception that redundancy improves sustainability; candidates may choose multi-region storage thinking it's good, but it increases energy use.

Full explanation →

457

Multi-Selectmedium

Which TWO actions are most aligned with responsible AI practices when deploying a model that makes decisions affecting individuals? (Choose 2)

Select 2 answers

A.Collect as much data as possible without quality checks

B.Continuously monitor the model for fairness metrics

C.Ensure the development team is homogeneous to avoid conflicts

D.Use the most complex model available for maximum accuracy

E.Provide meaningful explanations for model decisions

AnswersB, E

Fairness metrics can drift after deployment as data distributions shift, so continuous monitoring detects emerging bias against protected groups. This satisfies the responsible AI requirement for ongoing oversight of models making consequential decisions about individuals.

Why this answer

Option B is correct because responsible AI requires ongoing monitoring of deployed models for fairness metrics (such as demographic parity, equalized odds, or disparate impact) to detect and mitigate bias that can emerge or drift after deployment. Option E is correct because providing meaningful explanations for model decisions supports transparency and accountability, enabling affected individuals to understand and potentially contest decisions, which aligns with principles like explainability and due process. Option A is incorrect because indiscriminate data collection without quality checks introduces noise, bias, and privacy risks rather than improving responsible AI.

Option C is incorrect because a homogeneous development team increases the risk of blind spots and systemic bias, whereas diverse teams better surface fairness concerns. Option D is incorrect because maximizing model complexity for accuracy alone ignores interpretability, fairness, and other responsible AI trade-offs.

Full explanation →

458

MCQhard

A developer sets a low temperature value when calling a foundation model through Amazon Bedrock for a use case that extracts structured fields from invoices. A colleague argues that lowering temperature reduces hallucination and guarantees correct extraction. How should the developer respond?

A.Lowering temperature forces the model to retrieve the correct field values from the invoice image before generating text.

B.Lowering temperature disables the model's ability to generate any text that was not present in its training data.

C.Lowering temperature makes sampling more deterministic, which improves consistency but does not guarantee factual correctness or eliminate hallucination.

D.Lowering temperature guarantees the response will conform to the requested JSON schema without any prompt instructions.

AnswerC

Temperature scales the randomness of token sampling; a low value makes the model favor high-probability tokens so outputs become more repeatable. However, if the model's learned associations are wrong or the input is ambiguous, the deterministic output can still be incorrect. Consistency and accuracy are distinct properties, so the colleague's guarantee claim is unfounded.

Why this answer

Temperature governs the randomness of token sampling, so lowering it makes outputs more deterministic and repeatable. Determinism is not the same as correctness: a model can consistently produce the same wrong field value, and hallucinations can persist even at very low temperature. Accurate extraction therefore still depends on model capability, prompt design, and validation logic.

Exam trap

The trap here is equating low temperature with factual accuracy, when the parameter only makes sampling more deterministic and can consistently reproduce the same error.

Full explanation →

459

Multi-Selecteasy

A company wants to document their machine learning model's intended use, limitations, and ethical considerations. Which TWO practices should they adopt? (Choose two.)

Select 2 answers

A.Create a model card that describes the model's purpose, performance, and fairness metrics

B.Automatically retrain the model weekly

C.Use Amazon SageMaker Model Cards to version and share model documentation

D.Set up Amazon CloudWatch alarms to monitor model accuracy

E.Conduct A/B testing between different model versions

AnswersA, C

A model card documents intended use, limitations, performance and fairness metrics in one artefact, directly satisfying the requirement to record purpose, constraints and ethical considerations. It provides structured transparency for stakeholders reviewing the model's responsible AI posture.

Why this answer

Option A is correct because a model card is the standard artifact for documenting a model's intended use, limitations, performance, and fairness/ethical considerations, directly satisfying the requirement to document purpose and ethics. Option C is correct because Amazon SageMaker Model Cards is the AWS service purpose-built to create, version, and share this structured model documentation (including intended use, risk ratings, and evaluation results) with stakeholders. Option B is incorrect because weekly automatic retraining is an MLOps maintenance activity that improves freshness but does not document intended use, limitations, or ethics.

Option D is incorrect because CloudWatch alarms monitor operational metrics such as accuracy drift or latency and provide no documentation of purpose or ethical considerations. Option E is incorrect because A/B testing compares model versions for performance or business impact and does not produce the required documentation.

Exam trap

AIF-C01 often tests the distinction between governance/documentation practices (model cards) and operational practices (retraining, monitoring, A/B testing) — candidates pick monitoring tools because they sound responsible, but the question is specifically about documenting intent and ethics.

Full explanation →

460

Multi-Selectmedium

A company is using Amazon Bedrock to build a conversational agent. They want to ensure the agent maintains context across multiple turns in a conversation. Which TWO strategies should the developer implement? (Choose two.)

Select 2 answers

A.Use Amazon Bedrock's session management feature to automatically persist conversation state.

B.Include the entire conversation history in each prompt sent to the model.

C.Enable the model's memory parameter to retain information across API calls.

D.Use a single API call with a streaming response to keep the connection open and maintain context.

E.Store conversation history in an external database and retrieve relevant turns to include in the prompt.

AnswersB, E

Including the entire conversation history in each prompt allows the model to see previous exchanges and maintain context. This is a common technique for multi-turn conversations with foundation models, as they are stateless and do not remember previous interactions. However, it increases token usage and may hit context length limits, so it should be managed carefully.

Why this answer

To maintain context across multiple turns, developers must include conversation history in each prompt. This can be done by sending the full history or by storing history externally and retrieving relevant parts. Foundation models are stateless, so context must be explicitly provided.

Options suggesting built-in memory or session management are incorrect.

Exam trap

The trap here is assuming that Amazon Bedrock or the foundation models automatically maintain conversation state, when in fact they are stateless and require manual context management.

Full explanation →

461

MCQmedium

A media company wants to add a generative AI assistant to its internal knowledge portal. The assistant must answer employee questions using only the company's private policy documents, and it must cite the exact source passage for each answer. The team plans to use a foundation model hosted in Amazon Bedrock. Which approach should they implement to meet these requirements?

A.Raise the maximum token limit in the model invocation parameters so the entire policy library can be pasted into every prompt.

B.Fine-tune the foundation model on the full set of policy documents and deploy a custom model that answers from its updated weights.

C.Use Retrieval Augmented Generation by storing policy documents in a vector store and retrieving relevant passages to include in the prompt sent to the model.

D.Increase the model's temperature setting so it produces more varied and detailed answers from its pretrained knowledge.

AnswerC

Retrieval Augmented Generation retrieves the most relevant passages from an indexed knowledge base and injects them into the prompt, so the model conditions its answer on the company's own documents instead of only pretrained knowledge. Because the retrieved passages are known, the application can return them as citations. This directly satisfies both the grounding and citation requirements without retraining the model.

Why this answer

The requirements are grounding answers in private documents and citing exact source passages. Retrieval Augmented Generation satisfies both by indexing the policy corpus, retrieving the most relevant chunks at query time, and passing them into the prompt so the model can answer from supplied evidence. The application can then surface the retrieved passages as citations, keeping answers accurate and auditable without retraining the foundation model.

Exam trap

The trap here is assuming that fine-tuning or a larger context window can substitute for retrieval when the real requirement is traceable, up-to-date grounding in a private document set.

Full explanation →

462

MCQeasy

A developer is using Amazon Bedrock Agents to build a multi-step reasoning bot that can query a SQL database and summarize results. Which service should be integrated as the action group's Lambda function to execute SQL queries?

A.Amazon Aurora

B.Amazon DynamoDB

C.Amazon OpenSearch Serverless

D.Amazon S3

AnswerA

Amazon Aurora is a fully managed SQL-compatible relational database that can be queried using standard SQL, making it suitable for the Lambda action group to execute SQL queries.

Why this answer

Amazon Aurora is correct because it is a fully managed SQL-compatible relational database that can be queried using standard SQL. The Lambda function in the action group can use the boto3 library to execute SQL queries against Aurora, enabling the Bedrock Agent to retrieve and summarize data from a SQL database.

Exam trap

The trap here is that candidates may confuse DynamoDB's PartiQL support with full SQL, but PartiQL is a limited SQL-compatible query language and DynamoDB remains a NoSQL database not designed for complex SQL joins or aggregations typical of multi-step reasoning tasks.

How to eliminate wrong answers

Option B is wrong because Amazon DynamoDB is a NoSQL key-value and document database that does not support SQL queries; it uses its own API for data access. Option C is wrong because Amazon OpenSearch Serverless is a search and analytics engine, not a SQL database, and it uses OpenSearch query DSL, not SQL. Option D is wrong because Amazon S3 is an object storage service that does not support executing SQL queries directly; while it can be queried via Athena, the Lambda function itself cannot execute SQL against S3 without an intermediary service.

Full explanation →

463

MCQhard

A large enterprise has multiple teams deploying ML models on AWS. To ensure governance and accountability, they need to enforce that all models pass a fairness review before production deployment. Which SageMaker feature should they use to implement this approval workflow?

A.SageMaker Studio

B.SageMaker Experiments

C.SageMaker Model Monitor

D.SageMaker Model Registry

AnswerD

SageMaker Model Registry enforces the approval workflow: models are registered as versioned model packages, and a pending manual approval status blocks deployment until a fairness review is completed. This directly satisfies the enterprise's governance requirement that every model pass review before reaching production.

Why this answer

SageMaker Model Registry is the correct choice because it provides a centralized catalog for managing ML models, including versioning, approval status, and metadata. It supports approval workflows by allowing you to define model groups, set approval statuses (e.g., PendingApproval, Approved, Rejected), and integrate with CI/CD pipelines to enforce that only approved models are deployed to production.

Exam trap

The trap here is that candidates confuse SageMaker Model Registry with SageMaker Model Monitor, mistakenly thinking monitoring covers pre-deployment fairness checks, when in fact Model Monitor only handles post-deployment observability.

How to eliminate wrong answers

Option A is wrong because SageMaker Studio is an integrated development environment (IDE) for building, training, and deploying ML models; it does not natively enforce approval workflows or governance for model deployment. Option B is wrong because SageMaker Experiments is used for tracking and comparing ML training runs (e.g., hyperparameters, metrics), not for managing model approval or deployment governance. Option C is wrong because SageMaker Model Monitor is designed for detecting data drift and model quality degradation in production, not for pre-deployment approval workflows.

Full explanation →

464

Multi-Selectmedium

Which THREE are SageMaker built-in algorithms suitable for regression tasks?

Select 3 answers

A.Linear Learner

B.K-Means

C.PCA

D.DeepAR

E.XGBoost

AnswersA, D, E

Linear Learner trains a linear function to predict a continuous target, supporting both regression and classification. It is a SageMaker built-in algorithm explicitly designed for regression, satisfying the question's requirement for a suitable built-in option.

Why this answer

Linear Learner (A) is a SageMaker built-in algorithm that supports both classification and regression; in regression mode it learns a linear function minimizing a chosen loss (e.g., squared error) on continuous targets. DeepAR (D) is a built-in supervised time-series forecasting algorithm that predicts continuous numeric values, so it is used for regression-style forecasting tasks. XGBoost (E) is a built-in gradient-boosted trees algorithm whose objective can be set to reg:squarederror or reg:logistic, making it suitable for regression on tabular data.

K-Means (B) is an unsupervised clustering algorithm that groups data points and does not predict a continuous target, and PCA (C) is an unsupervised dimensionality-reduction algorithm, so neither belongs to regression tasks.

Exam trap

The AIF-C01 exam often tests the distinction between supervised and unsupervised algorithms, and the trap here is that candidates may confuse dimensionality reduction (PCA) or clustering (K-Means) with regression tasks, assuming any algorithm that processes numeric data can perform regression.

Full explanation →

465

MCQeasy

A data scientist needs to predict house prices based on features like square footage, number of bedrooms, and location. Which type of machine learning is most appropriate for this task?

A.Classification

B.Reinforcement learning

C.Regression

D.Clustering

AnswerC

Regression models a continuous numeric target, so it predicts a price value from features such as square footage, bedrooms and location. Classification would output discrete labels, which cannot represent the continuous house-price output the scenario requires.

Why this answer

Regression is the correct choice because the task involves predicting a continuous numerical value (house price) based on input features. Unlike classification, which predicts discrete categories, regression models the relationship between independent variables and a continuous target, making it ideal for price prediction.

Exam trap

AWS AI Practitioner exams often test the distinction between regression and classification by presenting a prediction task with a continuous output, where candidates mistakenly choose classification because they focus on the word 'predict' without recognizing the output type.

How to eliminate wrong answers

Option A is wrong because classification predicts discrete class labels (e.g., 'expensive' vs. 'cheap'), not a continuous value like price. Option B is wrong because reinforcement learning involves an agent learning through trial-and-error interactions with an environment to maximize cumulative reward, which is not applicable to static supervised prediction tasks. Option D is wrong because clustering is an unsupervised learning method that groups data points based on similarity without a target variable, whereas this task requires a labeled dataset with known house prices.

Full explanation →

466

MCQeasy

A data scientist wants to host a pre-trained model on Amazon SageMaker for real-time inference with minimal latency. Which approach should they use?

A.Run inference using AWS Lambda with the model packaged as a container

B.Use SageMaker batch transform

C.Create a SageMaker asynchronous inference endpoint

D.Deploy the model on a SageMaker real-time endpoint

AnswerD

A SageMaker real-time endpoint keeps model artefacts loaded on dedicated instances behind a persistent HTTPS endpoint, returning predictions synchronously with low latency. This satisfies the minimal-latency requirement for real-time inference, unlike batch transform or asynchronous inference, which introduce queuing and storage round-trips.

Why this answer

SageMaker real-time endpoints are designed for low-latency, synchronous inference. They keep the model loaded and ready to respond to individual requests, making them ideal for real-time applications where minimal latency is critical.

Exam trap

The AIF-C01 exam often tests the distinction between synchronous (real-time) and asynchronous inference patterns, and the trap here is that candidates may confuse 'asynchronous inference' with 'real-time' because both can handle requests, but only real-time endpoints guarantee minimal latency for individual predictions.

How to eliminate wrong answers

Option A is wrong because AWS Lambda has a maximum execution timeout of 15 minutes and limited memory (up to 10 GB), making it unsuitable for hosting large pre-trained models that require persistent, low-latency inference. Option B is wrong because SageMaker batch transform is an asynchronous, offline process for processing large datasets in batches, not for real-time inference with minimal latency. Option C is wrong because SageMaker asynchronous inference endpoints are designed for requests with large payloads and longer processing times, where immediate response is not required; they introduce queuing and processing delays that are incompatible with minimal latency requirements.

Full explanation →

467

MCQmedium

A retail bank has millions of unlabeled customer transaction records and wants to discover natural groupings of spending behavior without defining any categories in advance. The data science team plans to use an unsupervised learning approach. Which technique is designed for this goal?

A.Linear regression, which models the relationship between a continuous target and one or more input features.

B.Logistic regression, which estimates the probability that a record belongs to a particular class.

C.A decision tree classifier, which splits data using feature thresholds to predict a labeled category.

D.K-means clustering, which partitions records into groups so that members of a group are more similar to each other.

AnswerD

K-means is an unsupervised algorithm that groups unlabeled records into a chosen number of clusters based on feature similarity. It directly matches the bank's goal of discovering natural spending-behavior segments without predefined categories. The team can then profile each cluster, for example frequent travelers versus everyday local spenders, to inform marketing or risk decisions.

Why this answer

The bank has no predefined categories and wants to uncover structure in unlabeled data, which is the defining use case for unsupervised learning. K-means clustering groups similar records together so the team can inspect and name the resulting segments afterward, turning raw transaction behavior into actionable customer groupings.

Exam trap

The trap here is reaching for a familiar supervised algorithm such as regression or a classifier even though the data has no labels and no target to predict.

Full explanation →

468

Multi-Selecthard

A company uses Amazon Bedrock Agents for customer support. The agent needs to perform multi-step reasoning: first identify the customer's account, then check order status, and finally provide a resolution. Which THREE components must be configured to enable this workflow? (Select THREE.)

Select 3 answers

A.Agent orchestration to plan and execute multi-step reasoning

B.A Bedrock Knowledge Base with customer data

C.A Lambda function for each action group to execute the API calls

D.A Bedrock Guardrail to block out-of-scope questions

E.An action group for each API call (account lookup, order status)

AnswersA, C, E

Agent orchestration is the reasoning engine that decomposes the request, sequences the account lookup, order status check, and resolution steps, and decides which action groups to invoke. It directly enables the multi-step reasoning the stem requires.

Why this answer

Option A is correct because Amazon Bedrock Agents rely on the agent's orchestration (the ReAct-based prompt/orchestration layer) to decompose the user request into a plan and sequence the multi-step reasoning across account identification, order-status lookup, and resolution. Option C is correct because each action group is backed by a Lambda function that Bedrock invokes to actually execute the business logic/API calls (for example, calling the CRM or order-management API) and return results to the agent. Option E is correct because you must define an action group per API operation—such as one for account lookup and one for order status—with its OpenAPI schema so the agent knows the available actions, parameters, and when to invoke them.

Option B is not required: a Knowledge Base is for retrieval-augmented generation over documents, not for executing the transactional API calls this workflow needs. Option D is not required: a Guardrail adds safety/content filtering but does not enable the multi-step action execution workflow.

Exam trap

A common mistake is assuming a Knowledge Base is required for any data retrieval, but here the agent needs to call live APIs (account lookup, order status) via action groups and Lambda, not query a static knowledge base.

Full explanation →

469

MCQmedium

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

A.Fine-tune a base LLM on the policy documents monthly

B.Use a larger foundation model with a longer context window and paste all documents into each prompt

C.Train a custom model from scratch on the policy documents each month

D.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

AnswerD

RAG retrieves relevant passages from a vector store at query time and supplies them as context to the language model, so monthly document updates only require re-indexing rather than retraining. This satisfies the constraint that the team cannot afford model retraining each time policies change.

Why this answer

Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions based on the latest policy documents without retraining the model. By indexing the documents in a vector store and retrieving relevant chunks at inference time, the system can incorporate monthly updates simply by re-indexing the new documents, keeping the model static and avoiding costly retraining.

Exam trap

AWS often tests the misconception that fine-tuning or retraining is necessary for domain-specific knowledge, when in fact RAG provides a cost-effective, update-friendly alternative that avoids model modification.

How to eliminate wrong answers

Option A is wrong because fine-tuning a base LLM monthly on updated policy documents is expensive, time-consuming, and risks catastrophic forgetting, making it impractical for frequent updates. Option B is wrong because pasting all documents into each prompt would exceed the context window limits of even the largest models, leading to truncation, loss of information, and high token costs. Option C is wrong because training a custom model from scratch each month is prohibitively expensive and resource-intensive, requiring massive compute and data preparation, which is unnecessary when a pre-trained model can be used with RAG.

Full explanation →

470

Multi-Selecthard

A data engineer is using Amazon SageMaker Data Wrangler to prepare tabular data for ML. Which THREE data transformations are natively supported? (Choose three.)

Select 3 answers

A.One-hot encoding for categorical features

B.Audio feature extraction

C.Text vectorization using TF-IDF

D.Custom Python code via Pandas or Spark

E.Image resizing and normalization

AnswersA, C, D

Data Wrangler includes a built-in one-hot encoding transform that converts categorical columns into binary indicator features, selectable directly in the transformation list. It satisfies the native transformation requirement without custom code, unlike techniques requiring external libraries or manual scripting.

Why this answer

SageMaker Data Wrangler natively supports one-hot encoding as a categorical-encoding transform, which converts categorical features into binary indicator columns suitable for ML models, so option A is correct. It also provides a text vectorization transform using TF-IDF (and other text transforms like bag-of-words and n-gram) to convert text fields into numeric feature vectors, making option C correct. Data Wrangler additionally allows custom transformations through the Custom Transform node, where users can write their own Pandas or PySpark code, so option D is correct.

Options B and E are not native Data Wrangler tabular transforms: audio feature extraction and image resizing/normalization are handled by other AWS services or frameworks (for example, SageMaker Processing with librosa or image libraries), not by Data Wrangler's built-in transform list.

Exam trap

AWS often tests the distinction between natively supported transformations in SageMaker Data Wrangler versus those requiring external services or custom scripts, leading candidates to mistakenly select audio or image processing options that are not part of Data Wrangler's built-in capabilities.

Full explanation →

471

Multi-Selectmedium

Which AWS services can be used to build, train, and deploy custom machine learning models? (Choose two.)

Select 2 answers

A.Amazon Polly

B.Amazon Lex

C.AWS Deep Learning AMIs

D.Amazon Rekognition

E.Amazon SageMaker

AnswersC, E

AWS Deep Learning AMIs provide pre-configured Amazon EC2 images bundling frameworks such as TensorFlow and PyTorch with GPU drivers, satisfying the build-and-train requirement. They supply the compute environment for custom model development, though deployment needs a separate service. This addresses the training half of the stem directly.

Why this answer

AWS Deep Learning AMIs (C) are correct because they provide pre-configured Amazon EC2 images with popular deep learning frameworks (TensorFlow, PyTorch, MXNet) and GPU drivers, giving you the full environment needed to build and train custom models on your own infrastructure. Amazon SageMaker (E) is correct because it is a fully managed platform whose built-in capabilities (notebooks, training jobs, automatic model tuning, and hosting endpoints) let you build, train, and deploy custom ML models end to end. The other options are managed AI services with pre-trained models rather than tools for creating custom models: Amazon Polly (A) only converts text to lifelike speech, Amazon Lex (B) only builds conversational chatbots using NLU, and Amazon Rekognition (D) only provides pre-trained image and video analysis.

Exam trap

The trap here is that candidates confuse pre-built AI services (Polly, Lex, Rekognition) with platforms that allow custom model development, leading them to select services that only consume pre-trained models rather than build and train custom ones.

Full explanation →

472

MCQeasy

A media company uses Amazon Transcribe for automatic speech recognition. They discover the model has higher error rates for non-native English speakers. Which Responsible AI principle are they failing to uphold?

A.Fairness

B.Explainability

C.Robustness

D.Privacy

AnswerA

Fairness requires that a system perform equitably across demographic groups. Higher error rates for non-native English speakers show the transcription model delivers unequal accuracy for a particular group, breaching that principle rather than, say, transparency or privacy.

Why this answer

The model's higher error rates for non-native English speakers indicate a bias in the training data or model design that leads to disparate performance across demographic groups. This directly violates the Fairness principle of Responsible AI, which requires that AI systems treat all groups equitably and do not amplify existing societal biases. Amazon Transcribe's underlying acoustic and language models may have been trained predominantly on native English speech, causing systematic underperformance for non-native accents.

Exam trap

AWS often tests the distinction between Fairness and Robustness, where candidates mistakenly attribute performance disparities to a lack of robustness rather than recognizing it as a fairness issue stemming from biased training data.

How to eliminate wrong answers

Option B (Explainability) is wrong because the issue is not about the model's inability to explain its decisions, but about biased outcomes across different speaker groups. Option C (Robustness) is wrong because robustness concerns the system's resilience to adversarial inputs or noise, not its fairness across demographic groups. Option D (Privacy) is wrong because the problem does not involve unauthorized data access or exposure of personal information; it is a performance disparity unrelated to data protection.

Full explanation →

473

Multi-Selecthard

Which TWO practices help ensure responsible AI when deploying generative AI applications? (Select TWO.)

Select 2 answers

A.Deploy the model without any content filters to maximize creativity

B.Increase model size to improve accuracy at the expense of interpretability

C.Use only synthetic data for training to avoid privacy issues

D.Implement guardrails to filter harmful or inappropriate content

E.Monitor the model's outputs for bias and drift over time

AnswersD, E

Guardrails intercept prompts and responses, blocking harmful, inappropriate or policy-violating content before it reaches users. This enforces responsible AI by constraining generative output at runtime, satisfying the requirement to prevent unsafe material being surfaced in deployed applications.

Why this answer

Option D is correct because implementing guardrails—such as content filters, prompt shields, and safety classifiers—directly mitigates the risk of generative AI producing harmful, offensive, or inappropriate content, which is a core requirement of responsible AI deployment. Option E is correct because continuously monitoring model outputs for bias and drift ensures that the system remains fair and accurate as data distributions and user behavior change over time, enabling timely remediation. Option A is incorrect because removing content filters increases the risk of harmful outputs and violates responsible AI principles rather than ensuring them.

Option B is incorrect because increasing model size at the cost of interpretability reduces transparency and explainability, which are key responsible AI goals. Option C is incorrect because using only synthetic data does not by itself guarantee privacy or fairness and can introduce its own biases and quality issues.

Exam trap

AIF-C01 often tests whether candidates confuse 'model performance improvements' (size, accuracy) with 'responsible AI controls' (guardrails, monitoring, transparency), so options that sound like optimization tricks are the trap.

Full explanation →

474

MCQhard

A company uses Amazon Bedrock to generate code. They want to ensure the code follows security best practices and does not contain vulnerabilities. Which approach is most effective?

A.Implement a post-processing step using AWS WAF.

B.Use Amazon CodeGuru Security to review generated code.

C.Train a custom model on the company’s secure code.

D.Use a foundation model trained only on secure code.

AnswerB

Amazon CodeGuru Security applies static analysis with detector rules tuned to identify security vulnerabilities and deviations from coding best practices, directly satisfying the requirement that generated code be checked for vulnerabilities. Unlike generic content filters, it inspects code semantics, catching issues such as injection flaws and insecure data handling within the Bedrock output.

Why this answer

Amazon CodeGuru Security reviews code for security vulnerabilities and provides recommendations. Using a model trained on secure code may not be sufficient; WAF is for web traffic; training a custom model requires significant effort and may not catch all issues.

Full explanation →

475

Multi-Selecthard

A team is using Amazon Comprehend to analyze customer feedback for sentiment. They want to detect and mitigate potential bias against certain demographic groups. Which TWO approaches should they consider? (Choose TWO.)

Select 2 answers

A.Use AWS WAF to filter out biased comments.

B.Use AWS CloudTrail to audit API calls.

C.Use Amazon Rekognition to verify images.

D.Use SageMaker Clarify to compute bias metrics on the training data.

E.Use Comprehend custom classification with balanced training data across groups.

AnswersD, E

SageMaker Clarify computes bias metrics such as class imbalance and disparate impact across demographic groups in training data, exposing skew before it propagates into sentiment predictions. That satisfies the requirement to detect bias against specific groups systematically.

Why this answer

Option D is correct because SageMaker Clarify is the AWS service purpose-built to detect bias in datasets and models, computing metrics such as class imbalance (CI) and difference in proportions of labels (DPL) across demographic groups, which directly addresses measuring bias in the training data used for sentiment analysis. Option E is correct because Amazon Comprehend custom classification lets you train a custom model on your own labeled data, and ensuring that training data is balanced across demographic groups reduces the risk that the classifier learns and amplifies skewed associations, thereby mitigating bias in the resulting sentiment predictions. Option A is not appropriate because AWS WAF is a web application firewall that filters HTTP/S traffic against exploits like SQL injection and XSS, not a tool for detecting or mitigating bias in text sentiment.

Option B is not appropriate because AWS CloudTrail only records API activity for auditing and governance, and does not analyze or reduce bias. Option C is not appropriate because Amazon Rekognition performs image and video analysis (such as facial detection and moderation), which is unrelated to detecting bias in textual customer feedback sentiment.

Exam trap

The trap here is that candidates may confuse AWS WAF or CloudTrail as general-purpose bias detection tools, when in fact they serve entirely different security and auditing functions, while the correct approaches require specialized ML fairness services like SageMaker Clarify and balanced training data practices.

Full explanation →

476

MCQhard

A financial services company is using Amazon Bedrock to generate investment summaries. They must ensure that the model does not provide personalized financial advice, which is a regulatory requirement. Which AWS feature should they use to block the model from generating such advice?

A.Amazon Bedrock Guardrails with a denied topics policy.

B.Amazon Bedrock model evaluation with a custom metric for regulatory compliance.

C.Amazon Bedrock Agents with an action group that checks for financial advice keywords.

D.Amazon Bedrock Provisioned Throughput with a custom model that has been fine-tuned to avoid financial advice.

AnswerA

Amazon Bedrock Guardrails allows you to define denied topics, which are subjects the model should avoid. By specifying 'personalized financial advice' as a denied topic, the guardrail will block the model from generating content on that topic. This is the appropriate feature to enforce regulatory compliance by preventing certain types of responses.

Why this answer

To block the model from generating personalized financial advice, the company should use Amazon Bedrock Guardrails with a denied topics policy. This feature allows defining topics that the model must avoid, effectively preventing non-compliant responses. Other options like model evaluation or fine-tuning do not provide real-time blocking.

Exam trap

The trap here is thinking that fine-tuning or model evaluation can reliably block specific content, when in fact Guardrails with denied topics is the purpose-built feature for real-time filtering.

Full explanation →

477

Multi-Selectmedium

A healthcare startup is deploying an AI system to assist in diagnosing skin conditions from images. They want to follow the NIST AI Risk Management Framework. Which THREE practices should they implement?

Select 3 answers

A.Use Amazon SageMaker to continuously monitor model performance and retrain as needed

B.Document the model's intended use, performance, and limitations in a model card

C.Establish a human-in-the-loop process for uncertain diagnoses

D.Archive all training data in Amazon S3 Glacier for long-term retention

E.Deploy the model on AWS Lambda for serverless inference

AnswersA, B, C

Continuous monitoring with retraining directly supports the NIST AI RMF's MEASURE and MANAGE functions, which require ongoing performance tracking and risk treatment. For diagnostic imaging, drift detection and periodic retraining address the healthcare-specific risk of degraded accuracy, satisfying the framework's demand for iterative risk management.

Why this answer

Option A is correct because the NIST AI RMF's MEASURE and MANAGE functions require ongoing monitoring of model performance in production and retraining when drift or degradation occurs, and Amazon SageMaker Model Monitor plus retraining pipelines directly support this continuous risk-management lifecycle. Option B is correct because the MAP function calls for documenting intended purpose, performance characteristics, and known limitations; a model card is the standard artifact for this transparency and is especially critical in a clinical context. Option C is correct because the MANAGE function emphasizes human oversight of AI decisions, and a human-in-the-loop process for uncertain diagnoses ensures a qualified clinician reviews low-confidence outputs, reducing patient-safety risk.

Option D is not required by the NIST AI RMF; long-term archival in S3 Glacier is a data-retention choice that does not by itself address AI risk management, and the framework does not mandate any specific storage class. Option E is also not a framework requirement; AWS Lambda serverless inference is an architectural deployment decision, and the NIST AI RMF is technology-agnostic, so it neither prescribes nor favors Lambda over other inference options.

Exam trap

AIF-C01 often tests the tendency to select technically plausible but irrelevant AWS services (Glacier, Lambda) as answers to responsible AI questions — candidates must map options to NIST AI RMF functions rather than to general AWS capabilities.

Full explanation →

478

Multi-Selecthard

A company is developing an AI system for resume screening. They want to ensure fairness and reduce bias. Which THREE steps should they take in accordance with the NIST AI Risk Management Framework and AWS responsible AI principles?

Select 3 answers

A.Establish a governance process for regular auditing and human review of decisions

B.Measure bias metrics across demographic groups using SageMaker Clarify

C.Optimize the model solely for overall accuracy

D.Ensure the training data includes diverse representation across demographic groups

E.Remove all sensitive attributes (e.g., gender, race) from the dataset

AnswersA, B, D

Regular auditing and human review embed the NIST AI RMF's Govern and Manage functions into the screening workflow, catching biased outcomes that automated scoring alone would miss. This satisfies the fairness constraint by keeping a human accountable for adverse decisions affecting candidates.

Why this answer

The NIST AI Risk Management Framework emphasizes measuring bias, ensuring diverse data, and establishing governance. Removing sensitive features alone is insufficient due to proxy correlations. Relying solely on accuracy ignores fairness.

The three correct steps cover measurement, data diversity, and governance.

Full explanation →

479

Multi-Selecthard

A developer is using the Amazon Bedrock InvokeModel API with streaming enabled. They want to process partial results as they arrive. Which THREE steps are necessary to implement streaming correctly? (Select THREE.)

Select 3 answers

A.Call the standard InvokeModel API and wait for the entire JSON response before processing partial results.

B.Use the InvokeModelWithResponseStream API operation to request a streaming response.

C.Enable raw HTTP/2 callbacks on the Bedrock client to automatically invoke a callback for each partial result.

D.Process the response stream by iterating over event-stream records and decoding each chunk event's payload as it arrives.

E.Handle in-stream error events (for example, modelStreamErrorException) and consume the stream until completion to avoid truncated output.

AnswersB, D, E

Each chunk contains a piece of the output.

Why this answer

The correct options are B, D, and E. Option B is required because InvokeModelWithResponseStream is the specific Bedrock API operation that returns a stream of response chunks (via the responseStream event stream) rather than a single blocking response, which is exactly what the developer needs to process partial results as they arrive. Option D is required because the streaming response is delivered as an event stream of typed events such as chunk and modelStreamErrorException, so the client must iterate over the stream and decode each chunk's bytes (for example, base64-decoded JSON delta text) to consume partial output incrementally.

Option E is required because streaming responses can terminate with in-stream error events and the connection can drop mid-stream, so the implementation must handle these exceptions and read the stream to completion to avoid truncated output. Options A and C are not part of the necessary streaming implementation: they describe non-streaming or unrelated behaviors and are not marked correct.

Exam trap

The AWS AI Practitioner exam often tests the distinction between synchronous (waiting for full response) and asynchronous streaming (processing chunks as they arrive), and the trap here is that candidates may think raw HTTP callbacks or waiting for the complete response are valid streaming implementations.

Full explanation →

480

MCQmedium

A healthcare company needs to build a GenAI application that summarizes patient discharge notes. The compliance team requires that no medical record identifiers (MRIs) appear in the model's output. Which Bedrock Guardrails feature should be configured?

A.PII detection and redaction

B.Topic denial

C.Grounding check

D.Content filtering

AnswerA

PII detection and redaction identifies and masks sensitive data such as medical record identifiers in both prompts and model responses, directly satisfying the compliance requirement that no MRIs appear in output. Configuring it as a guardrail filter blocks or anonymises detected identifiers before the summarised discharge note reaches the user.

Why this answer

Bedrock Guardrails' sensitive information filters (PII detection and redaction) can identify and redact standard PII such as names, addresses, and Social Security numbers. Medical record identifiers (MRIs) are not a built-in PII category, so you must add a custom regex pattern under the sensitive information filter to match and redact the MRI format. This is the correct feature because it prevents MRIs from appearing in output.

Exam trap

AWS often tests the distinction between content filtering (which blocks toxic or unsafe content), topic denial (which blocks specific subjects), and sensitive information filters (which detect and redact data patterns). Candidates may assume built-in PII detection covers all identifiers, but custom regex patterns are required for domain-specific identifiers like medical record numbers.

How to eliminate wrong answers

Option B (Topic denial) is wrong because it restricts the model from discussing certain topics (e.g., medical advice) but does not detect or redact specific data patterns like MRIs. Option C (Grounding check) is wrong because it verifies that model responses are based on provided source documents, but it does not filter out embedded identifiers from the output. Option D (Content filtering) is wrong because it blocks harmful or offensive content (e.g., hate speech, violence) rather than structured data like medical record numbers.

Full explanation →

481

MCQeasy

A company uses Amazon Bedrock to generate product descriptions. They notice that the output sometimes contains incorrect information. What should they do to improve accuracy?

A.Increase the temperature parameter.

B.Implement Retrieval-Augmented Generation (RAG).

C.Use a larger foundation model.

D.Use AWS WAF to filter outputs.

AnswerB

Retrieval-Augmented Generation grounds responses in authoritative source documents retrieved at inference time, so the model cites real product data instead of relying solely on parametric memory. This directly reduces fabricated or incorrect details in generated descriptions.

Why this answer

Retrieval-Augmented Generation (RAG) enhances the accuracy of foundation model outputs by grounding the generation in authoritative, up-to-date external knowledge sources. Instead of relying solely on the model's parametric memory, RAG retrieves relevant documents or data from a vector database (e.g., Amazon OpenSearch Serverless) and injects them into the prompt context, reducing hallucinations and incorrect information in product descriptions.

Exam trap

AWS often tests the misconception that simply using a larger or more powerful model (Option C) is the universal fix for accuracy issues, when in fact the root cause of hallucinations is often a lack of grounded, retrievable context that RAG specifically addresses.

How to eliminate wrong answers

Option A is wrong because increasing the temperature parameter makes the model's output more random and creative, which would likely increase, not decrease, the frequency of incorrect information. Option C is wrong because using a larger foundation model does not inherently fix factual accuracy; larger models can still hallucinate or produce outdated information without access to current or domain-specific data. Option D is wrong because AWS WAF is a web application firewall that filters HTTP traffic for security threats (e.g., SQL injection, XSS) and has no mechanism to validate or correct the factual accuracy of generated text.

Full explanation →

482

MCQmedium

Refer to the exhibit. A company sets up a knowledge base for a customer support chatbot using Amazon Bedrock. Users report that the chatbot misses relevant details from long documents. Which change to the data source configuration would most likely improve retrieval?

A.Increase the chunk size in FIXED_SIZE chunking

B.Change chunking strategy to SEMANTIC

C.Add more documents to the S3 bucket

D.Change the embedding model to a larger one

AnswerB

Semantic chunking splits documents at natural meaning boundaries rather than fixed token counts, preserving coherent context within each chunk. This improves retrieval accuracy for long documents, since relevant details are less likely to be split across chunks and missed during embedding search.

Why this answer

Semantic chunking groups text based on meaning rather than fixed token counts, preserving the natural boundaries of concepts and paragraphs. This ensures that relevant details from long documents remain intact within a single chunk, improving retrieval accuracy for the chatbot.

Exam trap

AWS often tests the misconception that simply increasing chunk size or using a larger embedding model will improve retrieval, when the real bottleneck is the chunking strategy's ability to preserve semantic coherence.

How to eliminate wrong answers

Option A is wrong because increasing the chunk size in FIXED_SIZE chunking can cause chunks to contain multiple unrelated topics, diluting the semantic focus and making retrieval less precise. Option C is wrong because adding more documents to the S3 bucket does not address the core issue of poor chunking; it may even introduce more noise if the chunking strategy remains suboptimal. Option D is wrong because changing the embedding model to a larger one may improve representation quality but does not fix the fundamental problem of how documents are split; poorly chunked content will still lose relevant details regardless of the embedding model.

Full explanation →

483

Multi-Selecthard

A financial services company is preparing to train a machine learning model on customer transaction data. The data science team must address data quality concerns before training, because poor data directly harms model performance. Which TWO practices best improve the quality of the training data? (Choose two.)

Select 2 answers

A.Add more duplicate rows to increase the effective dataset size

B.Increase the learning rate so the model converges faster on noisy data

C.Handle missing values by imputing them or removing affected records after analysis

D.Remove all features that have any correlation with the target variable

E.Normalize or standardize numeric features so they share a comparable scale

AnswersC, E

Missing values can bias or destabilize training if left untreated. Analyzing the pattern of missingness and then imputing sensible values or removing affected records produces a cleaner dataset that better represents the population. This is a foundational data quality practice directly tied to the goal of improving model performance before training begins.

Why this answer

Improving data quality before training centers on correcting defects in the data itself. Handling missing values prevents bias and instability, while scaling numeric features ensures algorithms treat inputs comparably. Hyperparameter changes, row duplication, and deleting predictive features do not repair data defects and in several cases worsen results.

Exam trap

The trap here is treating model tuning knobs or dataset inflation as data quality fixes when quality work targets the data itself.

Full explanation →

484

MCQeasy

A company wants to build a generative AI application that can summarize customer support tickets. They need to ensure the model stays up-to-date with the latest product documentation without retraining. Which AWS service would best support this requirement?

A.Amazon Bedrock with Retrieval Augmented Generation (RAG)

B.Amazon Comprehend

C.Amazon Rekognition

D.Amazon SageMaker Ground Truth

AnswerA

Amazon Bedrock with RAG retrieves current product documentation from a knowledge base at inference time and injects relevant passages into the prompt, so summaries reflect the latest content without retraining. This directly satisfies the stem's constraint of staying up-to-date without retraining, since updating the underlying data store refreshes outputs immediately.

Why this answer

Amazon Bedrock with Retrieval Augmented Generation (RAG) is the correct choice because it allows the generative AI model to access and incorporate the latest product documentation from an external knowledge base without retraining. RAG works by retrieving relevant document chunks at inference time and injecting them into the model's context, ensuring responses reflect current information. This directly meets the requirement for staying up-to-date with evolving documentation while avoiding the cost and latency of full model retraining.

Exam trap

The trap here is that candidates may confuse Amazon Comprehend's text analysis capabilities (like summarization via extractive methods) with generative AI summarization, overlooking that Comprehend cannot incorporate external, dynamic knowledge sources without retraining.

How to eliminate wrong answers

Option B (Amazon Comprehend) is wrong because it is a natural language processing (NLP) service for extracting insights like sentiment, entities, and key phrases from text, but it does not provide generative AI summarization capabilities or a mechanism to dynamically incorporate updated documentation. Option C (Amazon Rekognition) is wrong because it is a computer vision service for analyzing images and videos, not for processing text-based customer support tickets or integrating with product documentation. Option D (Amazon SageMaker Ground Truth) is wrong because it is a data labeling service used to create training datasets for machine learning models, not a generative AI service that can summarize text or retrieve real-time information from external sources.

Full explanation →

485

MCQmedium

A machine learning engineer is preparing to fine-tune a foundation model for a specialized legal summarization task. The team has only a few thousand labeled examples and a limited budget. Which fine-tuning approach is most appropriate to adapt the model efficiently under these constraints?

A.Increase the context window size so the model can read entire legal contracts at once.

B.Perform full fine-tuning that updates every parameter in the network.

C.Use parameter-efficient fine-tuning that updates only a small subset of adapter weights.

D.Train the foundation model from scratch on the legal corpus.

AnswerC

Parameter-efficient fine-tuning methods such as LoRA update only a small number of added or selected parameters while freezing the base model. This dramatically reduces compute, memory, and storage needs, making it feasible with a few thousand examples and a limited budget. It adapts the model to legal summarization effectively without the cost of full retraining.

Why this answer

Parameter-efficient fine-tuning, such as LoRA, freezes the base model and trains only a small set of additional parameters. This cuts compute and memory requirements sharply, making adaptation feasible with limited data and budget. It still specializes the model for legal summarization, delivering the task adaptation the team needs without the expense of full retraining or training from scratch.

Exam trap

The trap here is equating fine-tuning with updating all model weights, overlooking that parameter-efficient methods achieve task adaptation at a fraction of the cost.

Full explanation →

486

Multi-Selecthard

An AWS AI practitioner is designing a document processing pipeline using Amazon Textract and Amazon Comprehend. The pipeline must extract text from PDFs, detect entities, and classify documents into categories (e.g., invoice, contract, report). Which THREE steps should be included in the pipeline? (Choose three.)

Select 3 answers

A.Use Amazon Comprehend to train a custom classifier for document type

B.Use Amazon Personalize to recommend document categories

C.Use Amazon Textract to extract text from the PDF

D.Use Amazon Comprehend to detect entities such as dates and amounts

E.Use Amazon Rekognition to analyze images in the PDF

AnswersA, C, D

Amazon Comprehend custom classification trains a model on labelled examples to assign documents to categories such as invoice, contract or report. This satisfies the stem's classification requirement, which entity detection alone cannot fulfil, since categorisation needs supervised training on document types.

Why this answer

Option C is correct because Amazon Textract is the AWS service purpose-built to extract text and structured data from PDFs and scanned documents via synchronous or asynchronous APIs such as DetectDocumentText and StartDocumentTextDetection. Option D is correct because Amazon Comprehend's entity detection (DetectEntities) natively identifies entities like DATE, QUANTITY, and other standard or custom entity types needed to pull dates and amounts from the extracted text. Option A is correct because Amazon Comprehend custom classification lets you train a custom classifier on labeled documents to categorize them into classes such as invoice, contract, or report, which is exactly the classification requirement.

Option B is wrong because Amazon Personalize is a recommendation service for personalized user experiences, not document categorization. Option E is wrong because Amazon Rekognition performs image and video analysis (objects, faces, labels) and does not extract document text or classify document types.

Exam trap

The trap here is confusing Amazon Personalize (a recommendation engine) with Amazon Comprehend (a natural language processing service) for classification tasks, and assuming Amazon Rekognition can extract text from PDFs when that is the role of Amazon Textract.

Full explanation →

487

MCQeasy

Which type of machine learning is used when a model learns to play a game by receiving rewards or penalties for its actions?

A.Reinforcement learning

B.Semi-supervised learning

C.Supervised learning

D.Unsupervised learning

AnswerA

Reinforcement learning trains an agent through interaction with an environment, using rewards and penalties to shape behaviour. The agent learns a policy that maximises cumulative reward, which is precisely how game-playing models improve their actions over time.

Why this answer

Reinforcement learning is correct because it involves an agent learning to make decisions by interacting with an environment, receiving rewards for desirable actions and penalties for undesirable ones. This trial-and-error approach, often modeled using Markov Decision Processes (MDPs), is specifically designed for sequential decision-making tasks like game playing, where the agent must maximize cumulative reward over time.

Exam trap

AWS often tests the misconception that reinforcement learning is a type of supervised learning because both involve feedback, but the key trap is that supervised learning uses immediate, correct labels while reinforcement learning uses delayed, evaluative rewards or penalties without explicit correct answers.

How to eliminate wrong answers

Option B is wrong because semi-supervised learning uses a small amount of labeled data with a large amount of unlabeled data to improve model accuracy, not reward/penalty signals from an environment. Option C is wrong because supervised learning relies on labeled input-output pairs to learn a mapping function, not on sequential actions and delayed rewards. Option D is wrong because unsupervised learning finds hidden patterns or structures in unlabeled data without any feedback signals, such as rewards or penalties.

Full explanation →

488

MCQhard

A company is deploying a generative AI model that produces text summaries of legal documents. To comply with responsible AI guidelines, which of the following is the most critical to ensure transparency?

A.Informing users that the summaries are generated by AI

B.Ensuring the model does not reflect biases from training data

C.Achieving high performance on summary quality metrics

D.Guaranteeing the summaries are factually accurate

AnswerA

Disclosing AI generation directly satisfies the transparency principle: users must know the summary is machine-generated, not human-authored, so they can weigh its reliability. This is the specific responsible AI mechanism for transparency, distinct from accuracy, fairness or accountability controls.

Why this answer

Transparency in responsible AI requires that users are clearly informed when they are interacting with AI-generated content, especially in high-stakes domains like legal document summarization. Option A directly addresses this by mandating disclosure, which builds trust and allows users to critically evaluate the output. Without such disclosure, users may mistakenly attribute human-level authority or accountability to the AI system.

Exam trap

AWS often tests the distinction between ethical principles (like transparency, fairness, and accountability) and candidates confuse 'mitigating bias' or 'ensuring accuracy' with the specific requirement of transparency, which is solely about disclosure and explainability.

How to eliminate wrong answers

Option B is wrong because while mitigating bias is an important ethical consideration, it is not the most critical factor for transparency; transparency focuses on disclosure and explainability, not on the model's internal fairness properties. Option C is wrong because high performance on summary quality metrics (e.g., ROUGE scores) does not inherently ensure transparency; a model can be highly accurate yet opaque in its decision-making process. Option D is wrong because guaranteeing factual accuracy is a matter of reliability and safety, not transparency; transparency is about informing users of the AI's involvement, not about the correctness of the output.

Full explanation →

489

Multi-Selectmedium

Which TWO techniques are commonly used to prevent overfitting in machine learning models? (Select TWO.)

Select 2 answers

A.Add more irrelevant features

B.Use cross-validation

C.Increase model complexity

D.Reduce the amount of training data

E.Use regularization

AnswersB, E

Cross-validation partitions the data into folds, training and validating on different subsets, so performance estimates reflect generalisation rather than memorisation of one split. It detects overfitting by exposing the gap between training and held-out fold error.

Why this answer

Option B (Use cross-validation) is correct because techniques like k-fold cross-validation estimate how well a model generalizes to unseen data by training and validating on different data splits, which helps detect and mitigate overfitting during model selection and hyperparameter tuning. Option E (Use regularization) is correct because methods such as L1 (Lasso) and L2 (Ridge) regularization add a penalty term to the loss function that constrains the magnitude of model weights, reducing variance and preventing the model from fitting noise in the training data. The unmarked options do not belong: adding more irrelevant features (A) increases dimensionality and noise, making overfitting worse; increasing model complexity (C) raises variance and typically worsens overfitting; and reducing the amount of training data (D) gives the model less information to learn generalizable patterns, also promoting overfitting.

Exam trap

AWS often tests the misconception that adding more data or features always helps model performance, when in fact irrelevant features or reducing training data can worsen overfitting, and candidates may incorrectly associate 'more complexity' with better generalization.

Full explanation →

490

MCQmedium

A financial services firm fine-tuned a generative AI model on Amazon SageMaker to summarize quarterly reports. The summaries often miss key financial metrics such as revenue and profit margins. The fine-tuning dataset contained full reports with summaries that included these metrics. The model appears to understand the reports but omits critical numbers. Which course of action would most likely improve the summaries?

A.Re-fine-tune using a carefully crafted dataset that includes explicit instructions to include key metrics and provides examples of correct summaries

B.Increase the maximum number of tokens in the summary

C.Switch to a different pre-trained model like Claude instead of the current one

D.Implement a post-processing Lambda function that extracts metrics from the original report and appends them to the summary

AnswerA

The dataset lacked supervision signalling that metrics matter, so the model learned to omit them. Re-fine-tuning with explicit instructions and exemplar summaries that retain revenue and profit figures supplies that signal, steering generation toward including critical numbers.

Why this answer

The model's failure to include key financial metrics despite having them in the training data indicates a misalignment between the training objective and the desired output. By re-fine-tuning with a dataset that explicitly instructs the model to include key metrics and provides correct examples, you directly teach the model to prioritize and extract those specific numerical values during summarization, addressing the root cause of omission.

Exam trap

AWS often tests the misconception that increasing output length or switching models will fix content omission, when the real solution lies in improving the fine-tuning data quality and instruction design.

How to eliminate wrong answers

Option B is wrong because increasing the maximum number of tokens does not force the model to include specific metrics; it only allows longer outputs, but the model may still choose to omit critical numbers if it hasn't learned to prioritize them. Option C is wrong because switching to a different pre-trained model like Claude does not guarantee that the new model will automatically include key metrics; the underlying issue is the fine-tuning data and instruction format, not the base model's architecture. Option D is wrong because implementing a post-processing Lambda function that extracts metrics and appends them is a brittle workaround that does not fix the model's behavior; it adds complexity and may produce inconsistent summaries, whereas the goal is to have the model generate complete summaries natively.

Full explanation →

491

MCQeasy

A company builds an AI system that generates medical diagnoses. To ensure patient safety and allow oversight, the company wants a human to review all high-risk predictions before they are acted upon. Which AWS service should they use?

A.AWS Lambda

B.Amazon SageMaker Ground Truth

C.Amazon Augmented AI (A2I)

D.AWS Step Functions

AnswerC

Amazon Augmented AI (A2I) provides built-in human review workflows, letting you route low-confidence or high-risk predictions to human reviewers before action. It directly satisfies the stem's requirement for human oversight of all high-risk medical diagnoses, integrating with Amazon Rekognition and custom models via the A2I API.

Why this answer

Amazon Augmented AI (A2I) is specifically designed to add human review workflows to machine learning predictions, including the ability to route high-risk or low-confidence predictions to human reviewers before action is taken. It integrates with Amazon SageMaker and other AWS services to create human review loops for sensitive use cases like medical diagnoses. This directly matches the requirement for human oversight of high-risk predictions.

Exam trap

AIF-C01 often tests the difference between Ground Truth (data labeling for training) and A2I (human review of predictions) — candidates confuse the two because both involve humans and ML.

How to eliminate wrong answers

Option A is wrong because AWS Lambda is a serverless compute service for running code in response to events; it does not provide human review workflows or oversight. Option B is wrong because SageMaker Ground Truth is used for labeling training data (building datasets), not for reviewing live model predictions in production. Option D is wrong because AWS Step Functions orchestrates workflows and can include human approval steps via callbacks, but it is a general-purpose orchestration service, not a purpose-built human-in-the-loop ML review service like A2I.

Full explanation →

492

MCQhard

An ML engineer wants to store training data in a format optimized for linear data scanning and columnar access in SageMaker. Which format is most appropriate?

A.JSON

B.Image (JPEG/PNG)

C.Parquet

D.CSV

AnswerC

Parquet is a columnar format, so SageMaker reads only the columns a query needs and scans them linearly, cutting I/O versus row-based formats. This matches the stated requirement for columnar access and efficient linear scanning of training data.

Why this answer

Parquet is a columnar storage format optimized for both linear data scanning and columnar access, making it ideal for training data in SageMaker. It reduces I/O by storing data by columns rather than rows, enabling efficient retrieval of specific features during model training.

Exam trap

AWS often tests the misconception that CSV is the most efficient format for training data, but Parquet's columnar storage and compression provide superior performance for linear scanning and columnar access in distributed ML pipelines.

How to eliminate wrong answers

Option A is wrong because JSON is a row-oriented text format that requires full parsing for columnar access, leading to high I/O overhead and slower linear scans. Option B is wrong because image formats like JPEG/PNG are binary and designed for visual data, not structured tabular data, and lack columnar access capabilities. Option D is wrong because CSV is a row-oriented text format that, while simple, requires scanning entire rows to access specific columns and lacks compression and schema optimization.

Full explanation →

493

MCQhard

A company uses Amazon SageMaker to host a model for fraud detection. The model must be re-evaluated for bias on a monthly basis. Which SageMaker feature can be used to detect bias in a trained model?

A.SageMaker Debugger

B.SageMaker Model Monitor

C.SageMaker Clarify

D.SageMaker Autopilot

AnswerC

SageMaker Clarify runs bias detection on trained models, computing metrics such as disparate impact across facets before and after training. It satisfies the monthly re-evaluation requirement by analysing the model directly, unlike monitoring or debugging tools.

Why this answer

SageMaker Clarify is the correct choice because it is specifically designed to detect bias in machine learning models and data. It provides built-in capabilities to evaluate bias metrics (e.g., difference in positive proportions, disparate impact) both before training (pre-training bias) and after training (post-training bias), making it suitable for the monthly re-evaluation requirement.

Exam trap

The trap here is that candidates confuse SageMaker Model Monitor (which monitors data drift) with bias detection, but Model Monitor does not evaluate model fairness or bias metrics.

How to eliminate wrong answers

Option A is wrong because SageMaker Debugger is used for monitoring training jobs in real time to detect issues like vanishing gradients or overfitting, not for bias detection. Option B is wrong because SageMaker Model Monitor focuses on detecting data drift and quality issues in deployed endpoints, not on evaluating model bias. Option D is wrong because SageMaker Autopilot automates the process of building, training, and tuning models, but it does not include built-in bias detection capabilities.

Full explanation →

494

MCQmedium

A company is using Amazon Bedrock's Converse API to build a conversational agent. They want the agent to maintain context across multiple turns. The agent should also be able to call external APIs to retrieve real-time data when needed. Which combination of features should they use?

A.Use the Converse API with streaming enabled and set max tokens to 4096

B.Use the InvokeModel API with a conversational prompt template and tool use configuration

C.Use the Converse API with system prompts and tool configuration

D.Use the InvokeModel API with a system prompt and store conversation history client-side

AnswerC

The Converse API maintains multi-turn conversation context through the messages parameter, while tool configuration supplies function definitions that let the model request external API calls for real-time data. System prompts set behaviour but do not provide tool invocation.

Why this answer

The Converse API is designed for multi-turn conversations and natively supports system prompts to set agent behavior and tool configuration for invoking external APIs. Option C correctly combines these features to maintain context across turns and enable real-time data retrieval via tool calls, which is exactly what the scenario requires.

Exam trap

AWS often tests the distinction between the Converse API (which handles multi-turn context and tool configuration natively) and the InvokeModel API (which requires manual context management and lacks built-in tool support), leading candidates to mistakenly choose a lower-level API when the higher-level one is more appropriate.

How to eliminate wrong answers

Option A is wrong because streaming and max tokens control response delivery and length, not context maintenance or external API invocation. Option B is wrong because the InvokeModel API is a lower-level API that does not natively manage conversation state or tool configuration; it requires manual handling of conversation history and tool integration. Option D is wrong because while storing history client-side can maintain context, the InvokeModel API lacks built-in tool configuration support, making external API calls more complex and less integrated than using the Converse API's tool configuration.

Full explanation →

495

MCQmedium

A government agency is building an AI assistant using Amazon Bedrock to help citizens understand eligibility rules for public benefits. The agency's legal team requires that the assistant never provide medical diagnoses, never discuss competitors' services, and refuse requests to draft legal documents. The agency also needs to log blocked interactions for periodic review. Which Amazon Bedrock feature should the team configure to enforce these restrictions and capture denials?

A.Amazon Bedrock Agents with action groups invoking Lambda functions

B.Amazon Bedrock Provisioned Throughput for dedicated model capacity

C.Amazon Bedrock model evaluation jobs comparing foundation models

D.Amazon Bedrock Guardrails with denied topics and content filters

AnswerD

Amazon Bedrock Guardrails lets teams define denied topics that block specific subject areas and content filters that screen harmful categories. When a prompt or response violates a guardrail, the interaction is blocked and can be logged for review. This directly enforces the agency's prohibitions on medical diagnoses, competitor discussion, and legal drafting while supporting the audit requirement.

Why this answer

Bedrock Guardrails is the runtime content governance feature that enforces denied topics and content filters, blocking disallowed prompts and responses while producing logs for review. Evaluation jobs compare models offline, Provisioned Throughput manages capacity, and Agents extend capabilities through Lambda actions. Only Guardrails applies policy at inference time and captures the blocked interactions the agency must audit.

Exam trap

The trap here is assuming that Amazon Bedrock Agents enforce topic restrictions, when agents only orchestrate actions and leave content policy to Guardrails.

Full explanation →

496

MCQmedium

A team is using Amazon SageMaker to deploy a real-time inference endpoint. The endpoint must be accessible only from a specific IP range and must automatically scale based on request volume. Which configuration meets these requirements?

A.Deploy the endpoint in a public subnet with a security group allowing the IP range.

B.Configure the endpoint with a VPC and attach a security group that allows inbound traffic from the IP range, and enable automatic scaling for the endpoint.

C.Deploy the endpoint with a VPC and use a Network Load Balancer with target group health checks.

D.Deploy the endpoint with an AWS WAF ACL to filter by IP and enable auto scaling for the endpoint.

AnswerB

VPC security group restricts by IP and automatic scaling handles demand.

Why this answer

Amazon SageMaker endpoints can be deployed within a VPC, allowing you to attach a security group that restricts inbound traffic to a specific IP range. Additionally, SageMaker supports automatic scaling for real-time endpoints using Application Auto Scaling, which adjusts the number of instances based on request volume metrics like InvocationsPerInstance.

Exam trap

The trap here is that candidates often confuse network-level access control (security groups in a VPC) with application-layer filtering (AWS WAF) or assume that a public subnet with a security group is sufficient, not realizing that SageMaker endpoints in a public subnet are still internet-facing and cannot be restricted to a specific IP range without a VPC.

How to eliminate wrong answers

Option A is wrong because deploying the endpoint in a public subnet exposes it to the internet, and a security group alone cannot restrict access to a specific IP range if the endpoint is publicly accessible; SageMaker endpoints in public subnets are not supported for IP-based restriction without a VPC. Option C is wrong because while a Network Load Balancer (NLB) can provide health checks and distribute traffic, SageMaker endpoints do not require an NLB for IP-based access control or scaling; the VPC and security group configuration already handles access control, and NLB is not a standard component for SageMaker endpoint deployment. Option D is wrong because AWS WAF is a web application firewall that operates at the application layer (HTTP/HTTPS) and is typically associated with API Gateway or CloudFront, not directly with SageMaker endpoints; SageMaker endpoints do not natively integrate with AWS WAF for IP filtering, and using WAF would not replace the need for VPC-based network controls.

Full explanation →

497

MCQhard

A team is fine-tuning an Amazon Titan Text model on a small dataset of legal documents. After fine-tuning, the model produces outputs that are factually incorrect and sometimes contradicts the training data. What is the most likely cause?

A.The batch size was too large, reducing gradient noise and causing mode collapse

B.The learning rate was set too high, causing the model to diverge

C.The model was fine-tuned for too many epochs on a small dataset, leading to overfitting

D.The base model was not designed for legal domain tasks

AnswerC

Excessive epochs on a small legal dataset let the model memorise noise rather than generalise, so it drifts from the training facts and emits contradictions. Overfitting is the mechanism producing the factually incorrect output described in the stem.

Why this answer

Fine-tuning a large language model like Amazon Titan Text on a very small dataset for too many epochs causes overfitting. The model memorizes the limited training examples instead of learning generalizable patterns, leading to outputs that are factually incorrect or contradictory when faced with inputs that differ slightly from the training data.

Exam trap

A common trap is to attribute poor fine-tuning results to a high learning rate, but in this scenario, overfitting on a small dataset with too many epochs is a more likely cause, especially when the model produces confident but incorrect outputs.

How to eliminate wrong answers

Option A is wrong because a large batch size reduces gradient noise and can stabilize training, not cause mode collapse; mode collapse is a GAN-specific issue, not typical for fine-tuning a Titan Text model. Option B is wrong because while a high learning rate can cause divergence, the described symptoms (factual errors and contradictions) are more characteristic of overfitting on a small dataset, not training instability. Option D is wrong because Amazon Titan Text is a general-purpose foundation model that can be fine-tuned for any domain, including legal; the base model's domain suitability is not the root cause when the dataset is small and training is excessive.

Full explanation →

498

MCQmedium

Refer to the exhibit. A developer has attached this IAM policy to their user. When trying to invoke the Anthropic Claude v2 model using the Bedrock runtime, they receive an AccessDeniedException. Which change to the policy would resolve the issue?

A.Add the bedrock:InvokeModelWithResponseStream action

B.Change the Action to bedrock:ListFoundationModels

C.Change the Resource to arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-v2

D.Remove the Resource element and set Effect to Deny

AnswerC

The policy's Resource currently points at a provisioned throughput ARN, which only authorises invoking models via a purchased throughput commitment. Switching to the foundation-model ARN grants access to on-demand invocation of Anthropic Claude v2, satisfying the Bedrock runtime call that triggered the AccessDeniedException.

Why this answer

The IAM policy's Resource element must specify the exact ARN of the foundation model being invoked. The original policy likely used a wildcard or incorrect ARN pattern. The correct ARN for Anthropic Claude v2 in us-east-1 is arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-v2, which grants permission to invoke that specific model via the Bedrock runtime.

Exam trap

AWS often tests the nuance that Bedrock foundation model ARNs use a double colon (::) and omit the account ID, leading candidates to choose a wildcard Resource or a malformed ARN.

How to eliminate wrong answers

Option A is wrong because adding the bedrock:InvokeModelWithResponseStream action would allow streaming responses, but the error is AccessDeniedException due to an incorrect Resource ARN, not a missing action. Option B is wrong because changing the Action to bedrock:ListFoundationModels would only allow listing models, not invoking them, which does not resolve the invoke failure. Option D is wrong because removing the Resource element and setting Effect to Deny would explicitly deny all actions, making the problem worse; the Effect should be Allow, and the Resource must be correctly specified.

Full explanation →

499

MCQeasy

An ML team notices that the training accuracy is 99% but validation accuracy is only 72%. Which concept best describes this situation?

A.Cross-validation error

B.Overfitting

C.Bias-variance tradeoff

D.Underfitting

AnswerB

Overfitting occurs when a model memorises training data, capturing noise rather than generalisable patterns, so training accuracy stays high while validation accuracy lags. The 99% versus 72% gap directly satisfies the stem's constraint: a large, persistent divergence between training and validation performance on unseen data.

Why this answer

The model achieves 99% accuracy on training data but only 72% on validation data, which is a classic symptom of overfitting. Overfitting occurs when the model learns noise and specific patterns in the training set too well, failing to generalize to unseen data. This is often caused by excessive model complexity, such as too many layers in a neural network or too deep a decision tree, relative to the amount of training data.

Exam trap

AWS often tests the distinction between overfitting and the bias-variance tradeoff, where candidates may confuse the tradeoff as the direct answer instead of recognizing that the specific symptom (high training accuracy, low validation accuracy) is the definition of overfitting.

How to eliminate wrong answers

Option A is wrong because cross-validation error is a technique used to estimate model performance by partitioning data into folds, not a description of the training-validation accuracy gap itself. Option C is wrong because the bias-variance tradeoff is a broader concept that explains the relationship between model complexity and error, but the specific situation of high training accuracy and low validation accuracy is directly overfitting, not just a tradeoff. Option D is wrong because underfitting would show low accuracy on both training and validation sets (e.g., both below 70%), not high training accuracy with a large gap.

Full explanation →

500

MCQmedium

A company uses Amazon Bedrock Agents to automate order processing. The agent needs to call an internal database to check inventory. Which component should be used to integrate the database query?

A.Knowledge base

B.Action group

C.Guardrail

D.Prompt template

AnswerB

Action groups define the OpenAPI schema and Lambda function that an Amazon Bedrock agent invokes to fulfil a task, letting it query the internal inventory database during order processing. They satisfy the stem's requirement to integrate the database call, since the agent cannot reach external systems without an action group.

Why this answer

Bedrock Agents use action groups to define external tools or APIs. Each action group has an associated AWS Lambda function that performs the actual work (e.g., querying a database). The agent can call the action group via natural language instructions.

Full explanation →

501

MCQmedium

A machine learning engineer notices that a SageMaker training job failed due to insufficient permissions to access a KMS-encrypted S3 bucket. The training job's IAM role has S3 access permissions. What should be done to resolve the issue?

A.Create a new KMS key and re-encrypt the data

B.Assign the SageMakerFullAccess policy to the role

C.Add a kms:Decrypt permission to the IAM role for the specific KMS key used to encrypt the S3 bucket

D.Change the S3 bucket's default encryption to S3-managed keys (SSE-S3)

AnswerC

S3 access alone cannot decrypt KMS-encrypted objects; the training job's IAM role also needs kms:Decrypt on the specific key. This satisfies the stem's constraint that S3 permissions already exist, isolating the missing key-level authorisation as the cause of failure.

Why this answer

The training job failed because the IAM role lacks permission to decrypt the KMS key used for S3 server-side encryption. Even with S3 access permissions, SageMaker cannot read encrypted objects without the kms:Decrypt action on the specific KMS key. Adding kms:Decrypt to the role's policy for that key resolves the issue.

Exam trap

The trap here is that candidates assume S3 permissions alone are sufficient, overlooking that KMS-encrypted objects require explicit kms:Decrypt permissions on the IAM role, not just S3 bucket policies or managed policies like SageMakerFullAccess.

How to eliminate wrong answers

Option A is wrong because creating a new KMS key and re-encrypting the data is unnecessary and disruptive; the existing key can be used if the IAM role is granted the proper decrypt permission. Option B is wrong because SageMakerFullAccess is an AWS managed policy that does not include KMS permissions for customer-managed keys; it only grants basic SageMaker and S3 access, so it would not resolve the KMS decryption failure. Option D is wrong because changing the bucket's default encryption to SSE-S3 removes KMS encryption, which may violate security or compliance requirements, and is an overreaction when a simple IAM permission update can fix the issue.

Full explanation →

502

Multi-Selecthard

A healthcare startup is deploying an AI model to assist with diagnosis. They want to comply with the EU AI Act, which classifies medical AI as high-risk. Which THREE requirements must they fulfill? (Choose three.)

Select 3 answers

A.Provide technical documentation and logs for traceability

B.Establish a risk management system throughout the AI system's lifecycle

C.Ensure human oversight to prevent or minimize risks

D.Minimize the amount of training data to only necessary data

E.Allow the model to make final decisions without human review

AnswersA, B, C

High-risk systems under the EU AI Act require technical documentation and automatic logging to enable traceability and post-market monitoring. This satisfies the stem's compliance constraint by letting regulators and deployers reconstruct the model's operation and verify conformity.

Why this answer

Under the EU AI Act, high-risk AI systems such as medical diagnostic AI must satisfy a set of mandatory requirements, and option A is correct because providers must supply technical documentation and maintain automatically generated logs that enable traceability and post-market monitoring of the system's operation. Option B is correct because Article 9 requires a risk management system that is established, implemented, documented, and maintained as a continuous iterative process throughout the entire lifecycle of the high-risk AI system. Option C is correct because high-risk systems must be designed to allow effective human oversight, enabling humans to prevent or minimize risks to health, safety, or fundamental rights, which is especially critical in a clinical diagnosis context.

Option D is not a stated EU AI Act requirement; the Act addresses data governance, quality, and representativeness rather than mandating minimization of training data to only necessary data. Option E is incorrect because allowing the model to make final decisions without human review directly contradicts the human oversight requirement for high-risk AI systems.

Exam trap

The trap is mixing GDPR principles (data minimization) into the AI Act requirements, and forgetting that human oversight is mandatory — candidates who pick 'model decides autonomously' fail the high-risk test.

Full explanation →

503

MCQeasy

Which of the following is a key advantage of using a pre-trained foundation model over training a model from scratch?

A.Reduces the amount of labeled data and compute resources needed

B.Eliminates the need for any fine-tuning

C.Guarantees perfect accuracy on domain-specific tasks

D.Allows the model to work offline without any cloud infrastructure

AnswerA

Pre-trained foundation models already encode general language patterns from massive corpora, so adaptation via fine-tuning or prompting needs far fewer labelled examples and far less compute than training from scratch, cutting cost and time to deployment.

Why this answer

Pre-trained foundation models, such as those based on transformer architectures, have already learned general language patterns from vast, diverse datasets during their initial training. This transfer learning approach drastically reduces the need for large amounts of labeled data and extensive compute resources when adapting the model to a specific downstream task, as only a relatively small fine-tuning step is required.

Exam trap

The trap here is that candidates may assume pre-trained models are 'plug-and-play' and require no further training, but the exam tests the understanding that fine-tuning is a critical step to adapt the model to specific tasks, not an optional or eliminable one.

How to eliminate wrong answers

Option B is wrong because fine-tuning is almost always necessary to adapt a pre-trained foundation model to a specific domain or task; the model's general knowledge must be specialized, and eliminating fine-tuning would result in poor performance on domain-specific tasks. Option C is wrong because no model, including pre-trained foundation models, can guarantee perfect accuracy on any task due to inherent biases, data limitations, and the stochastic nature of model outputs. Option D is wrong because pre-trained foundation models typically require cloud infrastructure or powerful local hardware for inference due to their large size and computational demands; they cannot simply work offline without any infrastructure.

Full explanation →

504

MCQmedium

A company wants to use a third-party foundation model from Amazon Bedrock but is concerned about data privacy because the model provider might store prompts and responses. How should they address this concern?

A.Enable Amazon Bedrock model invocation logging to capture all interactions

B.Review the third-party model provider's data handling policy and choose a model that does not retain data

C.Use Amazon Macie to scan prompts before they are sent

D.Use AWS KMS to encrypt the prompts and responses before sending to the model

AnswerB

Data retention is governed by each third-party provider's own terms, not by Amazon Bedrock itself. Reviewing the provider's data handling policy and selecting a model whose terms guarantee no retention directly addresses the privacy concern.

Why this answer

Each third-party model provider in Bedrock has its own data handling policies. Customers should review those policies and can choose models that do not store data or use Bedrock features like Guardrails to redact sensitive data. However, the direct action is to review the provider's policy and select a model that meets privacy requirements.

Full explanation →

505

MCQmedium

A company is using Amazon Bedrock to generate code snippets. They notice the model occasionally generates code that fails to compile. What is the most effective way to improve code quality without retraining?

A.Reduce the temperature parameter to 0 for deterministic output.

B.Increase the max token limit to allow the model to complete the code fully.

C.Fine-tune the model on a dataset of correct code snippets.

D.Use few-shot prompt engineering with correct code examples and formatting instructions.

AnswerD

Few-shot prompting supplies in-context examples of compilable code plus explicit formatting rules, steering the model's token distribution toward syntactically valid output without weight updates. This directly satisfies the no-retraining constraint, unlike fine-tuning, and targets the compilation failures observed in the stem.

Why this answer

Few-shot prompt engineering provides the model with explicit examples of correct code and formatting instructions, guiding it to generate syntactically valid code without modifying the underlying model. This approach leverages in-context learning to improve output quality by conditioning the model on desired patterns, which is more effective than parameter adjustments alone for addressing compilation errors.

Exam trap

A common misconception in AWS AI services is that adjusting hyperparameters like temperature or token limits can fix output quality issues, when in fact prompt engineering techniques like few-shot learning are the primary non-retraining methods for improving model behavior.

How to eliminate wrong answers

Option A is wrong because reducing temperature to 0 makes the model deterministic but does not inherently fix code compilation errors; it only reduces randomness, which may still produce incorrect or incomplete code. Option B is wrong because increasing the max token limit allows longer outputs but does not address the root cause of compilation failures, such as syntax errors or logical mistakes. Option C is wrong because fine-tuning requires retraining the model on a dataset of correct code snippets, which contradicts the question's constraint of 'without retraining' and is not a prompt-level solution.

Full explanation →

506

MCQeasy

A media company wants to automatically generate short video captions from uploaded audio files using a foundation model on AWS. The solution should transcribe speech and then produce concise captions. Which combination of AWS services should they use?

A.Amazon Kinesis Data Streams to ingest audio, then Amazon Bedrock to generate captions directly from the audio stream.

B.Amazon Rekognition to analyze the audio, then Amazon Bedrock to generate captions.

C.Amazon Polly to convert audio to text, then Amazon Comprehend to generate captions.

D.Amazon Transcribe to convert audio to text, then Amazon Bedrock to generate captions from the transcript.

AnswerD

Amazon Transcribe performs automatic speech recognition to convert audio into text, and Amazon Bedrock provides access to foundation models that can summarize or rewrite that transcript into concise captions. This two-step pipeline matches the requirement: transcription followed by generative caption creation. Both are managed AWS services, so the company avoids building and operating its own speech and language models.

Why this answer

The task requires two capabilities: speech-to-text and generative text creation. Amazon Transcribe provides accurate automatic speech recognition, and Amazon Bedrock supplies foundation models that turn the transcript into concise captions. Together they form a managed, scalable pipeline that meets the scenario without custom model development.

Exam trap

The trap here is confusing text-to-speech with speech-to-text, or assuming a foundation model can accept raw audio input directly.

Full explanation →

507

MCQeasy

A developer wants to test different prompt variations for a chatbot without making repeated API calls. Which Amazon Bedrock feature can help compare model responses?

A.Model evaluation on Amazon SageMaker

B.Amazon Bedrock Playground

C.AWS Security Token Service (STS)

D.Amazon CloudWatch Logs

AnswerB

The Playground provides an interactive console where multiple prompt variants can be run side by side and responses compared visually. This satisfies testing without repeated API calls, as experimentation happens within the interface rather than through programmatic invocation.

Why this answer

Amazon Bedrock Playground provides an interactive console interface where developers can test and compare different prompt variations, model configurations, and foundation models side-by-side without making repeated API calls. It allows real-time experimentation with parameters like temperature, top-p, and prompt engineering to observe how the model responds, making it the correct choice for this use case.

Exam trap

AWS certification exams often test the distinction between interactive experimentation tools (Playground) and backend evaluation or monitoring services (SageMaker, CloudWatch), leading candidates to mistakenly choose a service that handles model evaluation or logging rather than the one designed for real-time prompt comparison.

How to eliminate wrong answers

Option A is wrong because Model evaluation on Amazon SageMaker is a service for evaluating and monitoring the performance of machine learning models in production, not for interactively testing prompt variations in a chatbot context without API calls. Option C is wrong because AWS Security Token Service (STS) is used for generating temporary security credentials to manage access to AWS resources, not for comparing model responses or prompt testing. Option D is wrong because Amazon CloudWatch Logs is a monitoring and logging service for collecting and analyzing log data from AWS resources, not an interactive tool for testing prompt variations or comparing model outputs.

Full explanation →

508

Multi-Selecteasy

A developer is new to Amazon Bedrock and wants to understand the components of tokenization and context windows. Which TWO statements are correct? (Select TWO.)

Select 2 answers

A.A larger context window always produces better quality responses

B.Tokenization splits text only on whitespace

C.The context window determines the maximum number of tokens the model can process in a single request

D.The cost of a request is based on the number of tokens per second

E.Tokens can be words or subwords, depending on the tokenizer

AnswersC, E

The context window is the fixed token budget spanning prompt plus generated output for one request. Exceeding it forces truncation or rejection, so it directly determines the maximum tokens a model can process in a single request.

Why this answer

Option C is correct because the context window defines the hard limit on the total number of tokens (input plus output) that a foundation model can handle in one inference request on Amazon Bedrock, so exceeding it requires truncation or chunking. Option E is correct because modern tokenizers such as BPE or SentencePiece break text into tokens that may be whole words, subwords, punctuation, or characters, not fixed units. Option A is wrong because a larger context window only increases how much text can be supplied; response quality still depends on prompt design, model capability, and relevance of the included content.

Option B is wrong because tokenization does not split solely on whitespace; subword tokenizers split within words and handle punctuation and special characters. Option D is wrong because Bedrock pricing is based on the number of input and output tokens processed (per 1,000 tokens), not on tokens per second, which is a throughput metric.

Exam trap

In AWS exams, often test the misconception that a larger context window universally improves response quality, when in reality it can degrade performance due to the 'lost in the middle' effect, where models struggle to attend to relevant information in very long sequences.

Full explanation →

509

MCQeasy

A solutions architect is explaining the concept of a foundation model to a non-technical stakeholder. The stakeholder asks what distinguishes a foundation model from a traditional task-specific machine learning model. Which statement best describes a foundation model?

A.It is a large model pre-trained on broad data that can be adapted to many downstream tasks.

B.It is a model trained exclusively on a company's proprietary data for one specific use case.

C.It is a small model optimized to run only on edge devices with limited compute.

D.It is a rule-based system that follows hand-coded logic to produce deterministic outputs.

AnswerA

This is correct because foundation models are trained on vast, diverse datasets using self-supervised learning, giving them broad capabilities that transfer to many tasks through prompting or fine-tuning. This generality is the defining characteristic that separates them from narrow, single-purpose models, and it directly answers the stakeholder's question about what makes them distinct.

Why this answer

A foundation model is characterized by large-scale pre-training on broad data, which yields general capabilities that can be adapted to numerous downstream tasks via prompting, fine-tuning, or other techniques. This generality is what differentiates it from task-specific models built for a single narrow purpose, and it is the essence the architect needs to convey.

Exam trap

The trap here is assuming that a foundation model is defined by its size or deployment location rather than by its broad pre-training and adaptability across many tasks.

Full explanation →

510

MCQeasy

A hospital wants to build a system that automatically assigns a specialty department (for example, Cardiology, Neurology, or Orthopedics) to each free-text patient referral note. The hospital has a large archive of past referral notes that were already labeled by clinicians with the correct department. Which type of machine learning problem does this scenario describe?

A.Supervised learning, because the model learns from labeled historical examples to predict a categorical outcome.

B.Generative AI, because the model must create a new department name for every referral note it processes.

C.Unsupervised learning, because the model must discover hidden structure within the referral notes without guidance.

D.Reinforcement learning, because the model improves through rewards received after each department assignment.

AnswerA

The archive of referral notes already carries clinician-assigned department labels, so the model has ground-truth targets during training and predicts one category from a fixed set. That is the definition of supervised classification, where the algorithm learns a mapping from input text to a discrete label and can then generalize to new, unseen referral notes.

Why this answer

Because every historical referral note already has a clinician-assigned department, the training data is fully labeled and the target is one of several discrete categories. That combination defines supervised classification, in which a model learns the relationship between note text and department and then predicts the department for new notes.

Exam trap

The trap here is assuming that working with text automatically makes a problem unsupervised or generative, when the presence of labeled outcomes is what determines the learning type.

Full explanation →

511

MCQeasy

Which of the following is a key principle of responsible AI according to AWS?

A.Complexity

B.Speed

C.Profitability

D.Transparency

AnswerD

Transparency is a core responsible-AI principle, requiring that AI systems be explainable and their limitations, data usage and decision-making processes openly communicated. This satisfies AWS's emphasis on enabling users to understand how models reach outputs, supporting accountability and informed trust in deployed AI solutions.

Why this answer

Transparency is a core principle of responsible AI according to AWS, meaning that customers should be able to understand how an AI system makes decisions, including its inputs, outputs, and limitations. AWS emphasizes this through features like model cards in Amazon SageMaker, which document model performance, intended uses, and biases. This principle ensures that AI systems are explainable and auditable, building trust with users.

Exam trap

In AWS, responsible AI principles like transparency are ethical guidelines, not technical performance metrics. Candidates often confuse operational concepts (e.g., speed, complexity) with these ethical principles.

How to eliminate wrong answers

Option A is wrong because complexity is not a principle of responsible AI; in fact, AWS advocates for simplicity and clarity to avoid opaque 'black box' models that hinder explainability. Option B is wrong because speed, while important for performance, is not a responsible AI principle; AWS focuses on fairness, accountability, and transparency over raw processing velocity. Option C is wrong because profitability is a business goal, not an ethical guideline; AWS's responsible AI principles prioritize societal impact and user trust over financial gain.

Full explanation →

512

MCQmedium

A company uses Amazon SageMaker Ground Truth to label a dataset for a binary classifier. To reduce labeling bias, which workforce configuration is most appropriate?

A.Automatic labeling with Active Learning

B.Public workforce with no qualification

C.Private workforce of domain experts

D.Vendor managed workforce

AnswerC

A private workforce of domain experts satisfies the bias-reduction constraint by restricting labelling to vetted specialists with relevant subject knowledge, rather than anonymous crowd workers who may apply inconsistent or culturally skewed judgements. For binary classification, this yields more reliable ground-truth labels, though it costs more and scales slowly.

Why this answer

A private workforce of domain experts ensures that labeling is performed by individuals with deep knowledge of the data domain, which directly reduces labeling bias. Domain experts are less likely to misinterpret ambiguous data points and can apply consistent, informed judgment, thereby minimizing systematic errors that could skew the binary classifier's training data.

Exam trap

A common mistake is assuming that automated or crowd-sourced labeling is always less biased or more efficient. In AWS SageMaker Ground Truth, for specialized tasks, a private workforce of domain experts is critical to avoid introducing systematic labeling errors that degrade model fairness.

How to eliminate wrong answers

Option A is wrong because automatic labeling with Active Learning relies on the model's own predictions to label data, which can propagate and amplify existing biases present in the initial training data, rather than reducing labeling bias. Option B is wrong because a public workforce with no qualification introduces high variability in labeling quality and can increase bias due to lack of domain knowledge, inconsistent interpretation, and potential cultural or demographic biases among anonymous workers. Option D is wrong because a vendor managed workforce, while providing some quality control, typically uses generalist labelers who may lack the specific domain expertise needed to correctly label nuanced or specialized data, which can still introduce bias from misinterpretation.

Full explanation →

513

MCQhard

A financial services company is building a generative AI application using Amazon Bedrock. They need to ensure the model does not generate responses that violate regulatory topics, such as specific prohibited financial advice. Which Bedrock feature should they use to block entire topics?

A.Topic denial

B.PII detection

C.Content filtering

D.Grounding check

AnswerA

Topic denial lets you define specific topics that the model must not discuss, such as prohibited financial advice.

Why this answer

Amazon Bedrock's Topic denial feature allows administrators to define a list of prohibited topics. When the model attempts to generate content related to any of these defined topics, the request is blocked entirely. This is the correct mechanism for preventing the model from generating responses that fall under specific regulatory categories, such as prohibited financial advice.

Exam trap

AWS often tests the distinction between content filtering (which blocks based on harmfulness) and topic denial (which blocks based on specific subject matter), leading candidates to confuse the two when the question involves regulatory or compliance-based restrictions.

How to eliminate wrong answers

Option B (PII detection) is wrong because it is designed to identify and redact personally identifiable information (e.g., names, SSNs) from inputs and outputs, not to block entire topics or categories of content. Option C (Content filtering) is wrong because it applies adjustable thresholds to filter harmful or offensive content (e.g., hate, violence) based on severity levels, not to block specific regulatory topics like financial advice. Option D (Grounding check) is wrong because it verifies that model responses are factually supported by a provided reference source (e.g., a knowledge base), but it does not block entire topics; it only detects ungrounded or hallucinated claims.

Full explanation →

514

MCQmedium

A machine learning model achieves 99% accuracy on the training set but only 65% on the test set. Which phenomenon is the model experiencing?

A.Bias-variance tradeoff

B.Data leakage

C.Overfitting

D.Underfitting

AnswerC

A large gap between training accuracy (99%) and test accuracy (65%) is the signature of overfitting: the model has memorised training noise and fails to generalise. High variance across datasets, not high bias, explains the poor test performance.

Why this answer

The model's high accuracy on the training set (99%) but significantly lower accuracy on the test set (65%) indicates that it has memorized the training data, including noise and outliers, rather than learning generalizable patterns. This is the classic symptom of overfitting, where the model performs well on seen data but fails to generalize to unseen data.

Exam trap

Candidates often mistakenly select 'bias-variance tradeoff' because they associate the gap with variance, but the question explicitly asks for the phenomenon, not the underlying tradeoff concept.

How to eliminate wrong answers

Option A is wrong because bias-variance tradeoff is a conceptual framework describing the balance between underfitting (high bias) and overfitting (high variance), not a specific phenomenon of performance disparity between training and test sets. Option B is wrong because data leakage would typically cause both training and test accuracy to be artificially high or inconsistent in a different pattern (e.g., test accuracy higher than expected), not a large gap with training accuracy far exceeding test accuracy. Option D is wrong because underfitting would result in poor performance on both the training set and the test set (e.g., both below 70%), not high training accuracy with low test accuracy.

Full explanation →

515

MCQmedium

A company uses Amazon Comprehend to analyze customer sentiment. They discover the model performs poorly on text with slang from underrepresented groups. What is the most responsible action?

A.Restrict model use to only standard English

B.Remove slang from input before inference

C.Adjust the confidence threshold only for those groups

D.Collect more representative training data including slang

AnswerD

Collecting representative training data that includes slang from underrepresented groups addresses the root cause: the model's vocabulary and patterns were learned from unrepresentative text. This improves sentiment accuracy for those groups rather than masking the disparity.

Why this answer

The core principle of responsible AI requires that models be trained on data that is representative of the populations they serve. Amazon Comprehend's sentiment analysis is a supervised machine learning model; its poor performance on slang from underrepresented groups indicates a training data bias. Collecting more representative training data, including that slang, directly addresses the root cause by enabling the model to learn the linguistic patterns of those groups, improving fairness and accuracy without restricting access or masking the problem.

Exam trap

The trap here is that candidates may choose a quick-fix technical workaround (like removing slang or adjusting thresholds) instead of recognizing that the responsible AI approach requires addressing the root cause of bias through data representativeness, which is a core ethical and technical principle tested in the AIF-C01 exam.

How to eliminate wrong answers

Option A is wrong because restricting model use to only standard English is a discriminatory practice that excludes underrepresented groups, violating responsible AI principles of fairness and inclusivity; it does not fix the model's bias but rather avoids it. Option B is wrong because removing slang from input before inference is a data preprocessing workaround that does not address the underlying model bias; it discards valuable linguistic data and can alter the true sentiment of the text, leading to inaccurate results. Option C is wrong because adjusting the confidence threshold only for those groups is a post-hoc tuning that does not correct the model's learned bias; it may reduce false positives but does not improve the model's understanding of slang, and it introduces inconsistent decision boundaries that can be seen as unfair.

Full explanation →

516

MCQmedium

A financial services company uses Amazon Bedrock to power a customer-facing chatbot that answers questions about loan products. During a compliance review, auditors ask the team to demonstrate that the chatbot's responses are grounded in approved policy documents and that the model is not generating unsupported financial advice. Which AWS service or feature should the team use to trace each response back to the specific source passages used to generate it?

A.Amazon Bedrock Knowledge Bases with citations enabled

B.Amazon Bedrock Guardrails with contextual grounding checks

C.AWS CloudTrail with Bedrock data events logging

D.Amazon SageMaker Model Monitor with data quality baseline

AnswerA

Amazon Bedrock Knowledge Bases supports returning citations that map generated content back to the specific chunks retrieved from the underlying data source. This gives auditors a traceable link from each response to the approved policy document passages. Enabling citations in the RetrieveAndGenerate API response satisfies the requirement to demonstrate grounding and provenance for each chatbot answer.

Why this answer

The requirement is per-response provenance: showing which approved source passages produced each answer. Amazon Bedrock Knowledge Bases with citations enabled returns references to the retrieved chunks alongside the generated text, giving auditors a traceable map. Guardrails can detect ungrounded output but does not cite sources, while Model Monitor and CloudTrail address operational drift and API activity rather than content traceability.

Exam trap

The trap here is assuming that contextual grounding checks in Amazon Bedrock Guardrails produce source citations, when they only detect and filter ungrounded responses without returning passage-level references.

Full explanation →

517

MCQeasy

Which metric is most appropriate for evaluating a classification model when false positives are costly?

A.Precision

B.F1 score

C.Recall

D.Accuracy

AnswerA

Precision measures the proportion of positive predictions that are actually correct, so it directly penalises false positives. When false positives carry high cost, maximising precision minimises those costly incorrect positive classifications, unlike recall or accuracy.

Why this answer

Precision is the most appropriate metric when false positives are costly because it measures the proportion of true positive predictions among all positive predictions (TP / (TP + FP)). A high precision indicates that when the model predicts a positive class, it is very likely correct, minimizing the number of false positives. This directly aligns with the business requirement to avoid costly false alarms.

Exam trap

The AIF-C01 exam often tests the distinction between precision and recall by framing a cost scenario, and the trap here is that candidates confuse 'costly false positives' with 'costly false negatives' and incorrectly choose recall or F1 score without analyzing which error type is being penalized.

How to eliminate wrong answers

Option B (F1 score) is wrong because it is the harmonic mean of precision and recall, balancing both false positives and false negatives; it does not specifically penalize false positives more heavily. Option C (Recall) is wrong because it measures the proportion of actual positives correctly identified (TP / (TP + FN)), which is useful when false negatives are costly, not false positives. Option D (Accuracy) is wrong because it considers overall correct predictions (TP + TN) divided by total predictions, which can be misleading in imbalanced datasets and does not isolate the cost of false positives.

Full explanation →

518

Multi-Selecteasy

A developer is building an application that uses Amazon Bedrock to answer questions based on a large internal knowledge base. The knowledge base contains PDFs, Word documents, and web pages. Which TWO AWS services are commonly used together to implement a Retrieval-Augmented Generation (RAG) architecture on AWS? (Select TWO.)

Select 2 answers

A.Amazon Bedrock Knowledge Bases

B.Amazon SageMaker Ground Truth

C.AWS Glue

D.Amazon Athena

E.Amazon OpenSearch Serverless

AnswersA, E

Amazon Bedrock Knowledge Bases handles the ingestion, chunking and embedding of the PDFs, Word documents and web pages into a vector store, then retrieves the relevant passages at query time so the model answers from your internal corpus rather than its training data. This directly satisfies the RAG requirement for grounding responses in the knowledge base.

Why this answer

Amazon Bedrock Knowledge Bases (A) is correct because it is the managed RAG capability that ingests the PDFs, Word documents, and web pages, chunks and embeds them, and orchestrates retrieval so the foundation model can answer questions grounded in the internal knowledge base. Amazon OpenSearch Serverless (E) is correct because it is the commonly used vector store backing a Bedrock knowledge base, holding the embeddings and performing the vector similarity search that retrieves relevant passages at query time. Together they form the standard AWS RAG pattern: Bedrock Knowledge Bases for ingestion, embedding, and orchestration, and OpenSearch Serverless as the vector index for semantic retrieval.

Amazon SageMaker Ground Truth (B) is a data-labeling service for building training datasets, not a retrieval component. AWS Glue (C) is a serverless ETL/catalog service and Amazon Athena (D) is a serverless SQL query engine over S3; neither provides the vector similarity search or managed RAG orchestration this scenario requires.

Exam trap

The trap here is that candidates may confuse data preparation or query services (like AWS Glue or Athena) with the vector search and retrieval components essential for RAG, overlooking that Amazon OpenSearch Serverless provides the vector database capability while Bedrock Knowledge Bases orchestrates the ingestion and retrieval pipeline.

Full explanation →

519

MCQmedium

A company wants to use Amazon Bedrock to translate customer emails from English to Spanish. The emails contain occasional personal names and addresses. Which Guardrail configuration should be applied to protect customer privacy?

A.Grounding check

B.PII detection

C.Topic denial

D.Content filtering

AnswerB

PII detection identifies and redacts personal names and addresses in the email content before translation, satisfying the privacy requirement. It is the only guardrail filter that targets personally identifiable information specifically, rather than content safety or topic denial.

Why this answer

PII detection in Bedrock Guardrails can identify and redact personally identifiable information such as names and addresses. Content filtering, topic denial, and grounding check do not specifically target PII.

Full explanation →

520

MCQhard

A developer wants a foundation model to reliably return a structured record with fixed fields for downstream processing. The model currently returns free-form prose that breaks parsing. Which technique most directly improves output reliability for this use case?

A.Define a JSON schema and use constrained decoding or structured output features so the model can only emit tokens valid under that schema

B.Increase the model's temperature so it generates more varied field values

C.Shorten the prompt so the model has fewer instructions to follow

D.Add the phrase 'respond in JSON' to the user prompt and retry on parse failures

AnswerA

Constrained decoding restricts the token sampling space to sequences that satisfy the schema, guaranteeing syntactically valid structured output that downstream parsers can consume. This directly targets the parsing failures caused by free-form prose, unlike prompt wording changes that merely encourage format adherence.

Why this answer

Structured output reliability comes from constraining decoding to a schema, which makes invalid tokens impossible rather than merely unlikely. Prompt-based requests for JSON and retry loops reduce but do not eliminate malformed output, raising temperature increases variability, and shortening the prompt removes guidance without adding enforcement. Schema-constrained generation is the direct fix.

Exam trap

The trap here is believing that telling a model to produce JSON in the prompt is the same as guaranteeing valid JSON.

Full explanation →

521

Multi-Selecthard

A financial services firm is evaluating foundation models for a loan-summary assistant. They must consider both model characteristics and operational constraints. Which TWO factors most directly affect whether a candidate model can be deployed to meet their requirements? (Choose two.)

Select 2 answers

A.The popularity of the model among hobbyist developers on public forums

B.The model's maximum context window length relative to the size of the loan documents

C.The color scheme used in the model provider's marketing materials

D.Whether the model is available in the AWS Region where the firm must keep data

E.The number of parameters reported in the model's architecture diagram

AnswersB, D

Loan documents can be lengthy, and a model with a small context window cannot ingest the full text, forcing truncation or chunking that may drop critical clauses. Context length is therefore a hard feasibility constraint that determines whether the model can process the required input at all.

Why this answer

Deployability hinges on concrete constraints: whether the model can ingest the full loan document within its context window, and whether it is offered in the Region where data must remain. Parameter count, community popularity, and marketing presentation do not determine whether the model can satisfy the workload and compliance requirements.

Exam trap

The trap here is equating a model's general reputation or architecture size with its suitability for a specific regulated deployment.

Full explanation →

522

MCQmedium

A financial services company needs to use Amazon Bedrock to generate customer-facing content that must comply with strict regulatory guidelines. The company wants to minimize the risk of the model generating non-compliant content. Which technique should the company implement?

A.Use a higher temperature setting to make the model more conservative

B.Reduce the context window to limit the amount of input the model can see

C.Fine-tune the model on a dataset of compliant content only

D.Implement a guardrail that denies prohibited topics and enforces compliance rules

AnswerD

Guardrails in Amazon Bedrock apply policy-based filters that block prohibited topics and enforce compliance rules at inference time, directly satisfying the requirement to minimise non-compliant customer-facing output. Unlike prompt engineering or fine-tuning, guardrails provide deterministic, configurable denial of disallowed content regardless of the underlying foundation model's behaviour.

Why this answer

Amazon Bedrock Guardrails allow you to define denied topics, content filters, and compliance rules that are enforced at inference time, preventing the model from generating prohibited or non-compliant content. This is the most direct and reliable technique for regulatory compliance because it acts as a runtime safety layer, regardless of the underlying model or its training data.

Exam trap

AWS often tests the misconception that fine-tuning alone is sufficient for safety and compliance, when in reality guardrails are the recommended mechanism for enforcing runtime content policies in production.

How to eliminate wrong answers

Option A is wrong because increasing the temperature makes the model more random and creative, not more conservative; lower temperature values (closer to 0) produce more deterministic and conservative outputs. Option B is wrong because reducing the context window limits the input the model can see, but it does not prevent the model from generating non-compliant content based on the remaining input; it may even increase risk by removing context needed for accurate compliance. Option C is wrong because fine-tuning on compliant content only does not guarantee the model will never generate non-compliant content during inference, especially for edge cases or adversarial prompts; guardrails provide a hard enforcement layer that fine-tuning alone cannot match.

Full explanation →

523

MCQeasy

A data scientist is prototyping a text summarization application using Amazon Bedrock. They want to quickly test different prompts and models without writing code. Which AWS service or feature should they use?

A.AWS Cloud9

B.Amazon SageMaker Studio

C.AWS Lambda

D.Bedrock Playground

AnswerD

Bedrock Playground provides a console-based interface for interactively experimenting with prompts and switching between foundation models, letting the data scientist compare outputs without writing any code. It directly satisfies the rapid, code-free prototyping constraint in the stem.

Why this answer

Bedrock Playground is a no-code interface within Amazon Bedrock that allows users to experiment with different foundation models and prompts interactively. It is specifically designed for rapid prototyping without writing any code, making it the ideal choice for quickly testing text summarization prompts and models.

Exam trap

The trap here is that candidates may confuse SageMaker Studio (a full ML IDE) with a no-code testing environment, overlooking that Bedrock Playground is the dedicated service for quick, code-free model experimentation.

How to eliminate wrong answers

Option A is wrong because AWS Cloud9 is a cloud-based integrated development environment (IDE) for writing, running, and debugging code, not a no-code prompt testing tool. Option B is wrong because Amazon SageMaker Studio is a machine learning IDE that requires coding and setup for model experimentation, not a quick no-code playground for Bedrock models. Option C is wrong because AWS Lambda is a serverless compute service for running code in response to events, not an interactive interface for testing prompts without writing code.

Full explanation →

524

MCQmedium

A startup wants to generate high-quality images from text descriptions using Amazon Bedrock. They need to create realistic images of products for an e-commerce catalog. Which model provider should they choose?

A.Stability AI

B.Anthropic

C.Mistral AI

D.Cohere

AnswerA

Stability AI provides diffusion-based image generation models on Amazon Bedrock, such as Stable Diffusion, which convert text prompts into photorealistic product imagery. This directly satisfies the startup's requirement for high-quality, realistic e-commerce catalogue images generated from text descriptions, unlike providers offering only text or embedding models.

Why this answer

Stability AI is the correct choice because it specializes in image generation models, such as Stable Diffusion, which are designed to produce high-quality, photorealistic images from text prompts. Amazon Bedrock offers Stability AI's models for text-to-image tasks, making it ideal for generating realistic product images for an e-commerce catalog.

Exam trap

The trap here is that candidates may confuse general-purpose LLM providers (Anthropic, Mistral, Cohere) with specialized image generation models, assuming any AI model can generate images, but AWS Bedrock explicitly partitions providers by modality.

How to eliminate wrong answers

Option B (Anthropic) is wrong because Anthropic focuses on large language models (LLMs) like Claude, which are optimized for text generation, analysis, and conversation, not image generation. Option C (Mistral AI) is wrong because Mistral AI provides LLMs for natural language processing tasks, lacking any native image generation capabilities. Option D (Cohere) is wrong because Cohere specializes in embedding and retrieval-augmented generation (RAG) models for text, not image synthesis.

Full explanation →

525

Multi-Selecteasy

Which TWO are benefits of using Amazon SageMaker JumpStart for foundation models? (Choose 2)

Select 2 answers

A.Built-in fine-tuning scripts and notebooks

B.No coding required to fine-tune models

C.Automatic scaling without any configuration

D.Pre-trained foundation models available in the catalog

E.Free unlimited usage for all models

AnswersA, D

SageMaker JumpStart supplies pre-built fine-tuning scripts and notebooks, letting teams adapt foundation models without authoring training code from scratch. This directly satisfies the stem's benefit requirement by reducing implementation effort, since the notebooks run in SageMaker environments and expose hyperparameters for customisation.

Why this answer

Option A is correct because SageMaker JumpStart provides built-in fine-tuning scripts and example notebooks that let you adapt foundation models to your own data without writing the training pipeline from scratch. Option D is correct because JumpStart includes a catalog of pre-trained foundation models (e.g., from providers like Hugging Face, Meta, and AI21) that you can deploy or customize directly. Option B is not correct because fine-tuning still requires code or configuration, such as selecting hyperparameters and preparing datasets, so it is not fully no-code.

Option C is not correct because automatic scaling still requires configuring endpoint settings like instance counts and autoscaling policies. Option E is not correct because model usage in JumpStart is billed through SageMaker resources and is not free or unlimited.

Exam trap

AWS exams often test the misconception that 'no-code' solutions like SageMaker JumpStart eliminate all coding, when in reality they still require scripting for customization, and that built-in features like pre-trained models and scripts are distinct from automatic scaling or free usage.

Full explanation →

AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 451–525