AWS Certified AI Practitioner Practice Test (AIF-C01)

Retrieval-Augmented Generation (RAG) stores document embeddings in a vector database, retrieves the most relevant passages at query time, and injects them into the prompt sent to the foundation model. Because the model itself is not retrained, the process is faster and cheaper than full fine-tuning, pre-training, or collecting Reinforcement Learning from Human Feedback (RLHF) data. Fine-tuning or RLHF would require additional training cycles and do not automatically incorporate fresh content, while pre-training a new model would be far more costly and unnecessary for this scenario.

Diffusion models are trained to reverse a gradual noising process. At generation time, they start with pure random noise in the latent space. The model then performs a series of learned denoising steps that progressively remove noise, conditioning on the text prompt, until the noise is transformed into a coherent image. The other options do not describe the characteristic iterative denoising process: drawing pixels on a blank canvas is more like autoregressive image generation, retrieving and restyling an existing picture is nearest-neighbour plus style transfer, and a single decoding pass without iterations omits the fundamental step-by-step refinement that defines diffusion models.

When a prompt is specific and concise-for example, "List three advantages of Amazon Bedrock in one short sentence each" instead of "Tell me about Bedrock"-the model receives clear boundaries on scope and depth. This reduces ambiguity, limits the space for the model to digress, and therefore decreases the chance of verbose or off-topic content. Changing sampling parameters, adding unrelated examples, or removing context does not address the core problem that the instruction itself is too vague.

Amazon Comprehend is a fully managed natural language processing (NLP) service that can detect sentiment, extract key phrases, and identify entities in text with no ML expertise required. Amazon Translate focuses on language translation, Amazon Polly converts text to lifelike speech, and Amazon Rekognition analyzes images and videos, so none of those services provide built-in sentiment or key-phrase extraction.

Amazon Bedrock is designed so that customer content-such as prompts, responses, and embeddings-is encrypted in transit and at rest, and is not used by AWS or model providers to retrain foundation models. This built-in data-protection approach satisfies compliance needs for privacy-sensitive industries. The other statements describe capabilities that Bedrock does not currently provide: automatic GDPR audit report generation, on-premises model execution, or a global sub-100-millisecond latency SLA.

Simple, transparent models such as linear or logistic regression expose how each input contributes to the prediction, making them easy for auditors to understand. However, these models usually cannot capture complex patterns as well as deep neural networks, so accuracy can be lower. Deep neural networks often achieve higher predictive performance but are considered "black-box" and provide little built-in interpretability. The other options describe consequences that are not generally linked to transparency (such as needing more data, slower inference, or increased interpretability with deep models).

The dataset is imbalanced because one class (class A) overwhelmingly outnumbers the other. An imbalanced dataset can lead a model to favor the majority class and perform poorly on the minority class. A balanced dataset would have roughly equal representation of both classes, while the other listed terms do not describe class proportions.

Adding a human-in-the-loop review stage allows subject-matter experts to fact-check and approve each generated description before publication. This directly addresses the risk of inaccurate or hallucinated output. Raising the temperature usually increases creativity but can lower factual accuracy. Reducing the context window or turning off token streaming do not meaningfully improve correctness of the generated content.

The model is fitting the noise and details of the training data instead of learning the underlying patterns, so its predictions do not generalize to unseen data. This behavior reflects high variance, commonly called overfitting. High bias (underfitting) would show similar error on both training and test sets, data leakage would usually produce unrealistically good results on both sets, and a balanced model would maintain similar accuracy across datasets.

Choosing a smaller pre-trained model that already satisfies the application's accuracy requirements avoids the energy-intensive process of training a very large model from scratch. This reduces total compute time, electricity consumption, and associated carbon emissions. Retraining large models, running hardware 24/7 when not needed, or duplicating training jobs across Regions all increase resource use and environmental impact without adding sustainability benefits.

Human evaluation allows real people to judge nuanced, subjective aspects of a model's output-such as tone, helpfulness, or appropriateness-that automatic overlap-based metrics like BLEU, ROUGE, or BERTScore cannot fully capture. Automated metrics are fast and repeatable, but they focus on surface similarity to reference text and often correlate poorly with human perceptions of quality for conversational or creative tasks. Therefore, human review remains the best way to measure subjective response characteristics.

Service Organization Control (SOC) reports-SOC 1, SOC 2, and SOC 3-are produced under the American Institute of Certified Public Accountants (AICPA) framework. They provide third-party assurance about a service provider's internal controls, including security, availability, processing integrity, confidentiality, and privacy. AWS makes its SOC reports available to customers through AWS Artifact. ISO / IEC 27001, PCI DSS, and the GDPR are important standards and regulations, but they do not use the SOC report format.

Guardrails content filters include a predefined Misconduct category that detects prompts or responses that seek or provide information about engaging in criminal or fraudulent activity. Setting this category to Block causes the chatbot to refuse the request. Denied topics could work but require you to define and maintain custom topics, word filters only block exact words or phrases, and sensitive-information filters focus on PII rather than illegal-activity instructions.

Open source models publish their code, architecture details, and licensing. Because this information is openly accessible, teams can inspect how the model is built, audit the logic, and document limitations-key steps for transparency and explainability. Automatic global scaling, proprietary hidden data, or guaranteed top-tier accuracy are not inherent benefits of open source licensing.

Configuring PII redaction in Guardrails for Amazon Bedrock masks or removes personal data before it is logged or returned, directly addressing the privacy and security pillar of responsible AI. Bias analysis with SageMaker Clarify improves fairness, publishing a model card supports transparency, and word filters that block profanity improve content safety, but none of these measures specifically safeguard PII.

Creating an Amazon S3 interface VPC endpoint provides a private connection between resources in the VPC and Amazon S3 through AWS PrivateLink. Traffic routed through the endpoint remains on the AWS backbone, so it never traverses the public internet. Server-side encryption protects data at rest but does not control network paths. Amazon Macie helps discover sensitive data but also does not affect the network route. Granting SageMaker an IAM role controls authorization, not the physical connectivity path.

The described scenario is prompt poisoning. In a prompt-poisoning attack, adversaries inject harmful or misleading instructions into the data a foundation model later consumes (for example, documents indexed for RAG). When the model retrieves this tainted content, the malicious instructions become part of the prompt, leading the model to generate compromised outputs. Prompt hijacking instead refers to an attacker taking over an in-progress conversation, while jailbreaking tries to bypass a model's safety filters. Model drift is an unrelated performance change that happens gradually as data or requirements evolve, not from deliberate malicious instructions.

A customer managed policy lets you create a single reusable JSON policy document, then attach it to multiple IAM identities such as roles, users, or groups. This avoids duplication and simplifies updates, because you modify the policy in one place and the change propagates to every attached role. A role trust policy only controls who can assume the role; it does not grant S3 permissions. Service control policies apply at the AWS Organizations level and do not attach directly to individual roles. Adding an inline policy to every role is exactly what the team wants to avoid because it duplicates policy documents and is harder to maintain.

With steady, predictable traffic, Provisioned Throughput reserves dedicated model capacity for a fixed hourly fee. The reserved capacity delivers consistent low-latency responses and, when fully utilized, reduces the effective cost per request compared with pay-as-you-go On-Demand invocations. On-Demand remains more economical for sporadic or low traffic, but becomes more expensive per request at sustained high volume. Choosing a larger model or deploying to another Region does not directly lower per-request cost, and latency-based routing adds network overhead without addressing pricing.

A model is the mathematical representation produced by training algorithms on data. Once training finishes, this artifact captures learned patterns (for example, weights in a neural network) so it can accept new input and generate predictions or inferences. The algorithm itself, the raw training data, and the compute resources used are not the model; they are components of the pipeline that leads to creating the model.