AWS Certified AI Practitioner Practice Test (AIF-C01)

Tokenization breaks raw text into smaller units-such as words, sub-words, or characters-and maps each unit to a numeric token ID. This numerical sequence is what the model actually consumes. Splitting datasets, compressing parameters, or encrypting prompts are separate preprocessing or security tasks and are not part of tokenization itself.

Retrieval-Augmented Generation (RAG) works by first searching a designated knowledge source (such as the company's document repository), retrieving the most relevant passages, and then supplying those passages to the foundation model as additional context. Because the model receives up-to-date content at inference time, it can generate responses grounded in the company's current HR policies without the cost and complexity of pre-training or full fine-tuning.

Increasing temperature only affects randomness and does not introduce company-specific knowledge. Pre-training a new model from scratch and fully fine-tuning an existing model would embed the data, but both are far more expensive and time-consuming than RAG, and they require redeployment each time the documents change.

Guardrails for Amazon Bedrock offers a PII redaction capability that detects personally identifiable information in model outputs and masks or removes it before the response is returned. Denied topics policies block content about specific subjects but do not specifically target personal data. Content moderation severity thresholds rate harmful content, not PII exposure. A stop-sequence token simply truncates the model's text generation and is unrelated to PII protection.

Amazon Bedrock employs on-demand, pay-as-you-go pricing and fully manages the underlying GPU infrastructure. Because organizations are billed only for tokens processed during model inference or fine-tuning, they avoid both capital purchases and unused reserved capacity. The other options either impose fixed fees, require hardware ownership, or force multi-year commitments, so they do not provide the same cost-effective flexibility.

Prompt hijacking (often called prompt injection) occurs when a user supplies specially crafted input that convinces the model to ignore or overwrite the developer-provided system prompt. If successful, the attacker can make the chatbot reveal sensitive information, perform unintended actions, or deliver unapproved content. Data poisoning relates to corrupting the model's training data, not runtime instructions. Model underfitting and vanishing gradients are training issues, not security threats that arise from user input during inference. Therefore, preventing prompt hijacking-by techniques such as input sanitization, instruction hierarchy, or output validation-is the relevant concern in this scenario.

Embeddings map text (or other data) into high-dimensional numeric vectors in such a way that semantically similar items are located near one another in vector space. This representation enables semantic or "meaning-based" search because similarity can be computed with simple distance metrics. Tokenization merely splits text into tokens, chunking breaks long text into manageable pieces, and diffusion models are used for iterative data generation (e.g., images), none of which alone provide semantic vector representations for search.

Smaller, less complex models require fewer parameters and computational resources, allowing them to load into limited memory and generate outputs quickly on resource-constrained hardware. Larger or more complex models are designed to deliver extended context windows, broader multilingual coverage, or stronger emergent reasoning abilities, but they typically demand far more memory, power, and latency headroom than an edge device can provide.

AI/ML is valuable when a task involves uncovering patterns or making predictions that cannot be expressed with a fixed rule set. Converting prices between currencies uses a publicly posted exchange rate and a straightforward formula. A rule-based calculation is cheaper, fully accurate, and easier to maintain than building, training, and operating an ML model. The other scenarios involve classification, anomaly detection, or personalized recommendations-areas where ML can add measurable value.

Guardrails for Amazon Bedrock lets builders add safety controls to both user inputs and model outputs. Content filters are designed to detect and block or transform categories such as hate, harassment, sexual, or violent language. Because the requirement is to stop prompts with hateful or violent language before they reach the model, configuring a content filter on user inputs addresses the need.

• Denied topics policies block requests about broad subject areas (for example, medical advice) but are not category-based language filters.
• Word filters block specific custom terms rather than entire categories.
• PII redaction masks personally identifiable information and does not target hateful or violent language.

Temperature controls how randomly a model samples from its probability distribution when choosing the next token. A low value (for example, 0 or 0.1) narrows sampling to the most likely tokens, producing deterministic, repeatable answers. Raising temperature has the opposite effect, making responses more diverse. Changing max-tokens affects length, not randomness, and temperature is unrelated to training hyper-parameters such as learning rate.

Enabling S3 Versioning causes Amazon S3 to keep multiple versions of an object whenever it is overwritten or deleted. This allows earlier, uncorrupted versions to be restored, providing a straightforward safeguard for data integrity. Transfer Acceleration only speeds up uploads and downloads. S3 Object Lock places write-once-read-many (WORM) protections but does not by itself create previous copies when objects change. A Lifecycle rule that archives data to Glacier changes storage class but will not automatically preserve prior versions of modified objects.

Recommender systems analyze users' historical interactions-such as viewing and rating behaviour-to predict and surface items each user is likely to enjoy. This personalization objective is not addressed by fraud detection (focused on anomalous activity), speech recognition (turning audio into text), or image classification (labelling visual content).

Amazon SageMaker Model Cards generate a single, shareable document that captures the model's intended use, dataset sources, evaluation results, and responsible-AI considerations. Model Monitor tracks data and model drift, Clarify analyzes bias and explainability, and AWS Artifact provides compliance reports unrelated to model documentation.

Amazon S3 lifecycle configuration lets you define rules that automatically transition objects to different storage classes or expire (delete) them after a specified number of days. By creating a lifecycle rule that sets an expiration action at 1,095 days (3 years), the team can ensure datasets are deleted automatically, meeting the retention mandate. AWS Artifact provides compliance reports, CloudWatch Logs retention settings apply only to log groups, and AWS Config rules evaluate resource configurations but do not delete S3 objects.

Granting the SageMaker training job an IAM role that has least-privilege, read-only permission to the exact Amazon S3 prefixes keeps sensitive data access tightly controlled. Public buckets expose the data, embedding long-term keys in code violates security guidance, and removing encryption weakens protection; none of these align with proper data governance for fine-tuning.

A transparent or "white-box" model is valued because its internal parameters, features, and decision rules can be directly examined. This visibility allows practitioners and stakeholders to understand why the model produced a particular prediction, making it easier to audit for bias and ensure compliance with responsible-AI requirements. Models that instead maximize accuracy through complex, opaque structures, hide parameters, or rely on obfuscation are considered "black-box" approaches and do not provide the same level of interpretability.

The company's unmanaged outputs directly harmed the perceived reliability of its service. When customers lose confidence and stop using the product, the legal and business impact is categorized as a loss of customer trust. No intellectual-property dispute, data-sovereignty issue, or cost overrun is highlighted in the scenario, so those choices are not the primary risk demonstrated here.

The chatbot shows higher error rates for customers who use a regional dialect because that speech pattern is likely under-represented in the training data. This is an example of model bias that disproportionately affects a specific demographic group. The other options either contradict the observed accuracy difference, attribute the effect to an unrelated performance optimization, or incorrectly claim the underrepresented group experiences lower error.

RAG solutions first retrieve content relevant to the user's question and then pass that content to a foundation model for generation. Amazon Bedrock supports this by connecting to a vector database where embeddings of the source documents are stored. When a query arrives, the vector store performs similarity search, returns the most relevant passages, and those passages are appended to the prompt that Bedrock sends to the model. Fine-tuning the model, granting the model blanket S3 access, or only adjusting inference parameters does not provide retrieval capabilities. Therefore, adding a vector store such as an Amazon OpenSearch Service index for embeddings is required to implement RAG.

Stop sequences let you specify one or more character strings that signal the model to halt generation as soon as any of them is produced. Setting a stop sequence of "###END###" causes the model to terminate output when that exact string is emitted. Temperature adjusts randomness, Top-K limits the number of candidate tokens considered during sampling, and a maximum output token limit cuts off output only when the length threshold is reached-all of which do not stop generation based on a specific string.