AWS Certified AI Practitioner (AIF-C01) — All Questions

How does the Top K inference parameter influence text generation?

• a.It sets the AWS Region used for the request
• b.It restricts the next-token choice to the K most probable tokens at each step✓
• c.It updates the model's weights after each token is produced
• d.It encrypts the generated output before returning it

Top K limits sampling to the K highest-probability candidate tokens at each step, controlling diversity. It does not change weights, choose Regions, or encrypt output.

For a task that requires consistent, factual answers, how should you set the temperature parameter?

• a.As high as possible to maximize creativity
• b.Low (near zero) to reduce randomness and make outputs more deterministic✓
• c.Randomly for each request
• d.It has no effect on factual consistency

Lower temperature makes token selection more deterministic and focused, which suits factual, repeatable tasks. High temperature increases randomness and creativity, useful for brainstorming but riskier for facts.

What is the purpose of a 'stop sequence' inference parameter?

• a.It defines the persona the model should adopt
• b.It selects which foundation model to use
• c.It specifies text that, when generated, tells the model to stop producing further output✓
• d.It encrypts the prompt while it is in transit

A stop sequence marks where generation should end, useful for cutting off output cleanly at a delimiter. It does not set the model's persona, choose the model, or handle encryption.

What is a 'system prompt' in a conversational LLM application?

• a.High-level instructions that set the model's role, behavior, and constraints for the whole conversation✓
• b.The user's most recent chat message
• c.The billing summary generated for the session
• d.The list of GPUs assigned to the request

A system prompt establishes the assistant's role, tone, and rules that persist across the conversation, shaping how it responds to each user message. It is distinct from any single user turn.

Which Amazon Bedrock feature helps you compare foundation models and pick the best one for your use case using automatic metrics or human review?

• a.Amazon Bedrock Guardrails
• b.Amazon Bedrock Knowledge Bases for RAG
• c.Amazon Bedrock Agents
• d.Amazon Bedrock model evaluation✓

Bedrock model evaluation lets you assess and compare models with automatic metrics or human reviewers to choose the best fit. Guardrails enforce safety, Knowledge Bases enable RAG, and Agents orchestrate actions.

Which AWS service can serve as the vector store backing an Amazon Bedrock knowledge base?

• a.Amazon Polly
• b.Amazon OpenSearch Serverless✓
• c.Amazon Transcribe speech-to-text
• d.AWS CloudTrail

Amazon OpenSearch Serverless offers a vector engine that stores embeddings for similarity search behind a Bedrock knowledge base. Polly, Transcribe, and CloudTrail are unrelated to vector storage.

Besides OpenSearch, which AWS database can store vector embeddings for RAG by using the pgvector extension?

• a.Amazon Aurora PostgreSQL✓
• b.Amazon Redshift
• c.Amazon DynamoDB key-value tables
• d.Amazon Neptune

Amazon Aurora PostgreSQL supports the pgvector extension to store and query embeddings, making it a viable vector store for RAG. The other databases are not the pgvector-based option referenced here.

In a RAG pipeline on Amazon Bedrock, what is the role of an embeddings model such as Amazon Titan Text Embeddings?

• a.To generate the final natural-language answer for the user
• b.To convert documents and queries into numeric vectors so relevant chunks can be found by similarity✓
• c.To encrypt the knowledge base while it is at rest
• d.To provision the underlying servers for the pipeline

The embeddings model turns text into vectors so the system can retrieve the most semantically relevant chunks. A separate generation model then writes the final answer using those retrieved chunks.

Why are large documents split into smaller 'chunks' before being stored in a knowledge base for RAG?

• a.To permanently delete the original documents
• b.To translate the documents into another language
• c.To increase the model's temperature
• d.So the most relevant passages can be retrieved and fit within the model's context window✓

Chunking lets retrieval return just the passages that matter and keeps them within the context window, improving relevance and efficiency. It does not delete, translate, or change inference parameters.

In Amazon Bedrock Agents, what does an 'action group' define?

• a.The billing alerts configured for the account
• b.The list of users allowed to chat with the agent
• c.The APIs or AWS Lambda functions the agent can call to take actions on the user's behalf✓
• d.The encryption keys used to protect the model

Action groups tell a Bedrock agent which operations, via API schemas and Lambda functions, it may invoke to complete tasks, enabling multi-step automation. They are about capabilities, not billing or keys.

What does fine-tuning a foundation model on Amazon Bedrock require that prompt engineering does not?

• a.A physically larger monitor
• b.A curated training dataset plus additional compute time and cost✓
• c.Turning off all of the safety filters
• d.A dedicated public internet connection for each and every inference call

Fine-tuning trains on a prepared dataset and consumes extra compute, time, and money, unlike simply editing a prompt. It is worth it when prompting cannot reliably deliver the needed behavior.

Which Amazon Bedrock customization method further trains a model on large amounts of your unlabeled domain text to adapt its general knowledge?

• a.Zero-shot prompting
• b.Continued pre-training✓
• c.Provisioned Throughput capacity reservation
• d.Guardrails

Continued pre-training uses large unlabeled domain corpora to adapt a model's broad knowledge. Fine-tuning instead uses labeled examples; Provisioned Throughput is a capacity option and Guardrails handle safety.

A production application needs guaranteed, consistent inference capacity for a specific Amazon Bedrock model. Which option fits?

• a.Provisioned Throughput✓
• b.Spot capacity
• c.On-demand inference pricing
• d.An EC2 savings plan

Provisioned Throughput reserves dedicated model capacity for steady, predictable performance, often required for large or latency-sensitive workloads. On-demand bills per token but does not reserve capacity.

Which Amazon Bedrock pricing approach charges per token processed with no long-term commitment, suiting variable workloads?

• a.Provisioned Throughput
• b.Reserved Instances
• c.A one-year commitment plan
• d.On-demand inference✓

On-demand inference bills for the tokens you actually process, with no commitment, ideal for spiky or exploratory use. Provisioned Throughput instead reserves capacity for steady, high-volume needs.

Which is a prompt-engineering best practice for getting better, more relevant outputs?

• a.Be specific and provide clear context, constraints, and the desired output format✓
• b.Keep the prompt as vague as possible
• c.Never give the model any examples
• d.Always request the longest possible answer

Clear, specific prompts with context, constraints, and a target format reduce ambiguity and steer the model to the intended result. Vague prompts leave too much room for off-target responses.

Telling a model 'You are an experienced tax advisor...' before giving it the task is an example of what?

• a.Fine-tuning
• b.Role (persona) prompting✓
• c.Data augmentation technique
• d.Vector search

Assigning the model a role or persona in the prompt guides its tone, framing, and domain focus. It shapes the response through wording only and does not change the model's weights like fine-tuning.

How does grounding a model's answers in retrieved source documents (RAG) improve reliability?

• a.It makes the model train itself continuously
• b.It removes the need to review outputs at all
• c.It reduces hallucination by basing responses on real, relevant source content✓
• d.It lowers the token price to zero

By supplying verified, relevant passages at query time, RAG anchors answers in actual sources, cutting down on fabricated content. It does not eliminate the value of human review or change pricing.

Where in the Amazon Bedrock console can you quickly test prompts against different foundation models before building an application?

• a.AWS CloudTrail
• b.The Bedrock playgrounds (text, chat, and image)✓
• c.AWS Config
• d.Amazon Macie

Bedrock playgrounds let you experiment with prompts and compare model responses interactively before you build. CloudTrail, Macie, and Config are security and governance tools, not prompt-testing surfaces.

An application needs fast, low-cost responses for a simple text-classification task. What is a reasonable model choice?

• a.A smaller, less expensive foundation model that still meets the accuracy requirement✓
• b.Always the largest available model regardless of cost
• c.A text-to-image generation model
• d.A speech-to-text model

Matching model size to task keeps latency and cost down; a smaller model that meets accuracy targets is often the smart choice for simple tasks. The biggest model is not automatically the best fit, and image or speech models suit other tasks.

Where do you typically store the source documents that an Amazon Bedrock knowledge base ingests?

• a.Inside an IAM policy
• b.In an Amazon S3 bucket connected as the knowledge base data source✓
• c.Inside a CloudTrail log
• d.Inside the model's weights

Knowledge bases commonly ingest documents from an Amazon S3 data source, then embed and index them for retrieval. IAM policies grant permissions and CloudTrail logs activity; neither stores the documents.

In which situation is prompt engineering usually sufficient, making fine-tuning unnecessary?

• a.When you must permanently embed many thousands of proprietary domain facts directly into the model itself
• b.When the task can be handled by clearly instructing the model and optionally giving a few examples✓
• c.When the model must adopt an entirely new language it never saw in training
• d.When only a full retraining could possibly meet the requirement

If a well-crafted prompt (optionally with a few examples) reliably produces the needed output, fine-tuning's added cost and effort are unnecessary. Deeply embedding large bodies of knowledge is where customization earns its cost.

An app lets users upload a photo and then ask questions about it in text. Which foundation-model capability does this require?

• a.A multimodal model that accepts both image and text inputs✓
• b.A text-only embeddings model
• c.A speech synthesis model
• d.A time-series forecasting model

Answering text questions about an uploaded image requires a multimodal model that understands both images and text. Text-only, speech, or forecasting models cannot interpret the image content.

When evaluating a foundation model for a summarization use case, which criterion is most relevant?

• a.The color of the console theme
• b.The number of AWS Regions available
• c.Whether the summaries are accurate, relevant, and faithful to the source✓
• d.The alphabetical order of the model provider's name

For summarization, quality means faithful, accurate, and relevant summaries, so those are the criteria to evaluate. Superficial factors like console theme or provider name are irrelevant to output quality.

What is a 'prompt template'?

• a.A reusable prompt structure with placeholders that are filled in with specific inputs at runtime✓
• b.A trained copy of the model's weights
• c.A vector database index
• d.An IAM permission boundary

A prompt template is a reusable skeleton with variables (placeholders) that get populated per request, ensuring consistent, maintainable prompts. It is not model weights, a database, or an IAM construct.

Which capabilities does Amazon Bedrock Guardrails provide for a generative AI application?

• a.Filtering harmful content, denying specified topics, and redacting sensitive information such as PII✓
• b.Automatic model training and deployment
• c.Provisioning virtual servers and networking
• d.Generating images from text prompts

Bedrock Guardrails enforce responsible-use policies by filtering harmful content, blocking disallowed topics, and redacting sensitive data like PII. They are a safety layer, not a training, infrastructure, or image-generation tool.

Which set best represents core dimensions of responsible AI as emphasized by AWS?

• a.Latency, throughput, and storage cost
• b.Alphabetical ordering, file size, and color
• c.Marketing reach, sales, and branding
• d.Fairness, explainability, robustness, privacy and security, transparency, and governance✓

AWS frames responsible AI around dimensions like fairness, explainability, robustness, privacy and security, transparency, governance, veracity, and controllability. Latency and cost are operational concerns, not responsible-AI dimensions.

Which Amazon SageMaker capability continuously watches a deployed model for data quality issues and prediction drift?

• a.Amazon SageMaker Canvas
• b.Amazon SageMaker Model Monitor✓
• c.Amazon SageMaker JumpStart
• d.Amazon SageMaker Data Wrangler for data preparation

SageMaker Model Monitor tracks deployed models for drift and data-quality problems, alerting when behavior degrades. Canvas builds no-code models, JumpStart offers prebuilt models, and Data Wrangler prepares data.

Which Amazon SageMaker feature documents a model's intended use, training details, and risk ratings to support governance?

• a.Amazon SageMaker Model Monitor
• b.Amazon SageMaker Model Cards✓
• c.Amazon SageMaker Ground Truth
• d.Amazon SageMaker Canvas

SageMaker Model Cards capture a model's purpose, data, performance, and risk information in one place for governance and accountability. Model Monitor watches drift, Ground Truth labels data, and Canvas builds models.

A dataset used to train a face-analysis model contains mostly images of one demographic. What problem does this most directly create?

• a.Higher inference latency
• b.Sampling (data) bias that can cause unfair, unequal performance across groups✓
• c.Reduced storage cost
• d.A larger context window

An unrepresentative sample introduces bias, so the model tends to perform worse for underrepresented groups, an unfairness problem. It is a data-representation issue, not one of latency, storage, or context size.

Automatically removing customers' personal identifiers from data used with an AI system primarily supports which responsible-AI dimension?

• a.Privacy and data protection✓
• b.Latency optimization
• c.Cost reduction
• d.Model explainability and interpretability

Stripping or redacting personal identifiers protects individuals' privacy and limits exposure of sensitive data, supporting the privacy and security dimension. It is distinct from explainability or performance tuning.

A team tests whether a model still behaves safely and correctly when given unusual or adversarial inputs. Which responsible-AI dimension does this address?

• a.Robustness✓
• b.Transparency
• c.Explainability
• d.Cost efficiency

Robustness is about reliable, safe behavior even under noisy, unexpected, or adversarial inputs. Transparency and explainability concern openness and understanding decisions, which is a different concern.

Grounding outputs in trusted sources and checking them for factual accuracy chiefly supports which responsible-AI goal?

• a.Veracity and truthfulness, reducing hallucination✓
• b.Higher throughput
• c.Lower latency
• d.Smaller model size

Ensuring outputs are truthful and well-grounded targets veracity, directly countering hallucination. Latency, throughput, and model size are performance attributes, not truthfulness safeguards.

Why is human review important for content generated by an LLM in a public-facing product?

• a.It guarantees the model will never be updated again
• b.Humans can process data faster than any computer can
• c.It removes the need to secure the underlying data
• d.Generative models can hallucinate or produce biased content, so human oversight guards against harm✓

Because LLMs can fabricate facts or reflect bias, human review provides a safety check and accountability before content reaches the public. It complements, rather than replaces, security and monitoring.

Which action most directly helps reduce bias in a model's outcomes?

• a.Increasing the temperature parameter
• b.Training on data that is representative and balanced across relevant groups✓
• c.Deleting all monitoring logs
• d.Choosing a Region closer to users

Bias often stems from unrepresentative data, so curating balanced, representative training data is a primary mitigation. Inference settings, log deletion, and Region choice do not address the root cause of bias.

Disclosing to users that they are interacting with an AI system and explaining its limitations supports which principle?

• a.Latency
• b.Scalability
• c.Cost control and budgeting
• d.Transparency✓

Being open about the system's AI nature, capabilities, and limits so users can judge it appropriately is transparency. The other options are operational or financial attributes, not disclosure principles.

Designing an AI system so operators can guide, override, or shut it down when needed reflects which responsible-AI dimension?

• a.Controllability✓
• b.Fairness
• c.Explainability and interpretability
• d.Portability

Controllability is the ability to steer, correct, or stop an AI system's behavior, an important safeguard. Explainability is about understanding decisions and fairness about equitable outcomes, which are separate dimensions.

Which AWS service makes it easy to add human review of ML predictions when confidence is low or the stakes are high?

• a.Amazon Translate language translation service
• b.Amazon Polly
• c.Amazon Augmented AI (Amazon A2I)✓
• d.Amazon Forecast

Amazon A2I builds human-review workflows into ML applications, routing uncertain or high-stakes predictions to people. Polly, Translate, and Forecast are unrelated application services.

Which AWS service creates and manages the cryptographic keys used to encrypt data for AI and ML workloads?

• a.Amazon Macie
• b.AWS CloudTrail
• c.AWS Key Management Service (AWS KMS)✓
• d.Amazon Comprehend

AWS KMS creates, stores, and controls the encryption keys used to protect data at rest across services. Macie discovers sensitive data, CloudTrail logs activity, and Comprehend analyzes text.

What is the recommended way to give an application running on AWS temporary permissions to call Amazon Bedrock?

• a.Assign an IAM role, rather than embedding long-term access keys in the application✓
• b.Hard-code the root account credentials into the app
• c.Share one access key across the whole team
• d.Disable authentication for the service entirely

IAM roles grant temporary, automatically rotated credentials to applications, avoiding risky long-term keys. Root credentials, shared keys, and disabled auth all dramatically increase security risk.

Under Amazon Bedrock's data privacy terms, what happens to the prompts and data a customer sends to a base foundation model?

• a.They are published publicly for transparency
• b.They are sold to advertisers
• c.They are not used to train the base foundation models and are not shared with model providers✓
• d.They are automatically shared with all other customers

Amazon Bedrock does not use your prompts or completions to train the base models, and your data is not shared with third-party model providers, supporting confidentiality. It stays within your control.

How can a company access Amazon Bedrock or SageMaker without its traffic traversing the public internet?

• a.By disabling all encryption on the network connection
• b.By making the S3 buckets public
• c.By emailing the data to AWS
• d.By using VPC endpoints powered by AWS PrivateLink✓

VPC endpoints via AWS PrivateLink provide private connectivity to AWS services, keeping traffic off the public internet. Disabling encryption or making buckets public would reduce security, not improve privacy.

To create an audit record of every prompt sent to and response received from Amazon Bedrock, which feature would you enable?

• a.Model invocation logging to Amazon CloudWatch and/or Amazon S3✓
• b.Amazon Polly speech output
• c.Amazon Rekognition content moderation
• d.AWS Budgets spending alerts

Bedrock model invocation logging captures request and response details to CloudWatch Logs and/or S3 for auditing and troubleshooting. Polly, Rekognition, and Budgets serve unrelated purposes.

Where can a customer download AWS compliance reports, such as SOC and ISO certifications, to support governance?

• a.Amazon Kendra
• b.AWS Artifact✓
• c.Amazon Polly
• d.Amazon Textract

AWS Artifact is the self-service portal for AWS compliance reports and agreements (SOC, ISO, PCI, and more). The other services search content, synthesize speech, or extract document text; none provide compliance artifacts.

Which service collects metrics, logs, and alarms to monitor the operational health and usage of AI workloads?

• a.Amazon Macie for sensitive-data discovery
• b.AWS Artifact
• c.Amazon Translate
• d.Amazon CloudWatch✓

Amazon CloudWatch gathers metrics and logs and can trigger alarms, giving operational visibility into AI and other workloads. Macie finds sensitive data, Artifact provides compliance docs, and Translate converts languages.

What is the recommended AWS service for securely storing and automatically rotating an API key used by an AI application?

• a.AWS Secrets Manager✓
• b.Amazon S3 with public access enabled
• c.A plaintext file in the code repository
• d.Amazon Rekognition

AWS Secrets Manager securely stores credentials and can rotate them automatically, keeping secrets out of code. Public S3 or plaintext-in-repo would expose secrets, and Rekognition is a vision service.

Which AWS service uses machine learning and threat intelligence to continuously detect malicious activity across your AWS accounts?

• a.Amazon Comprehend
• b.Amazon GuardDuty✓
• c.Amazon Polly
• d.Amazon Lex

Amazon GuardDuty continuously analyzes account activity and network data to detect threats and anomalies. Comprehend analyzes text, Polly synthesizes speech, and Lex builds chatbots; none are threat-detection services.

A company must keep its data within a specific country to meet compliance rules. Which choice most directly supports this for an AI workload?

• a.Increasing the model's temperature
• b.Making the data public
• c.Selecting an AWS Region located in that country to process and store the data✓
• d.Turning off CloudTrail logging

Choosing an AWS Region in the required country keeps data processing and storage within that jurisdiction, supporting data-residency requirements. Inference settings and logging toggles do not address residency.

How can you keep an Amazon SageMaker training job isolated from the public internet?

• a.By running it under the root account
• b.By making the training bucket public
• c.By increasing the instance size
• d.By running the job within a VPC with network isolation enabled✓

Running SageMaker jobs inside a VPC with network isolation prevents outbound internet access, tightening security for sensitive data. Instance size, root usage, or public buckets do not provide network isolation.

Which AWS service helps continuously collect evidence and map it to frameworks to simplify audits of AI/ML and other workloads?

• a.AWS Audit Manager✓
• b.Amazon Polly
• c.Amazon Rekognition
• d.Amazon Translate

AWS Audit Manager automates evidence collection and maps it to compliance frameworks, streamlining audits. Polly, Rekognition, and Translate are AI application services, not audit-automation tools.

Across many AWS accounts, how can a company centrally enforce guardrails such as restricting which AI services or Regions may be used?

• a.By sharing a single common IAM user across all of the member accounts
• b.By applying AWS Organizations service control policies (SCPs)✓
• c.By disabling CloudTrail everywhere
• d.By making every S3 bucket public

AWS Organizations SCPs set permission guardrails across accounts, so you can centrally restrict services, actions, or Regions. Sharing users, disabling logging, or public buckets would weaken security and governance.