AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 976–1000

Option B is correct because prompt engineering allows the company to embed brand guidelines directly into the instruction given to the model, effectively steering the output away from prohibited keywords and tones. Option C is correct because Amazon Bedrock Guardrails provides a managed, policy-based mechanism to define custom deny topics and content filters that can block or mask unwanted content at inference time, enforcing brand guidelines without manual intervention.

The first step is to identify and understand the bias through measurement. SageMaker Clarify can compute fairness metrics like false negative rate difference to quantify the disparity. Retraining with fairness constraints or reweighting are mitigation steps that come after measurement.

SageMaker Role Manager helps administrators create least-privilege IAM roles, but it does not restrict which models can be used. Model access in SageMaker is controlled via IAM policies on the CreateTrainingJob and CreateModel APIs, combined with resource-level permissions.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Option A is correct because Amazon Lex provides the DeleteTranscript API, which allows the deletion of user conversation transcripts stored by Lex. This directly supports GDPR compliance by enabling the removal of personal data from conversation logs. The API targets the specific transcripts associated with a user, ensuring that their data is permanently deleted.

The bucket policy includes a condition 'aws:SourceAccount' that restricts access to requests originating from a specific AWS account. If the Rekognition service is being called from a different account (e.g., the Lambda function's IAM role is in a different account than the one specified in the condition), Rekognition will be denied read access to the S3 bucket. This condition is often misconfigured, causing cross-account access failures even when the principal is allowed.

Increasing validation loss while training loss decreases is a classic sign of overfitting. Early stopping halts training when validation loss starts to rise, preventing overfitting. The other options either worsen overfitting or do not address the root cause.

Guardrails provide both topic denial and content filtering. Topic denial blocks specific subjects, and content filtering handles offensive language.

Batch inference processes multiple records asynchronously, which is cost-effective for large volumes. On-demand and provisioned throughput are for real-time use cases.

Bedrock Agents can orchestrate translation steps; prompt engineering can instruct the model to translate. Knowledge Bases and Guardrails are not translation solutions. Human agents are not automated.

Amazon Comprehend can extract entities and perform topic modeling to understand the ticket content, while Amazon Kendra provides a searchable knowledge base that can be updated with new categories dynamically. Together they allow routing based on semantic understanding without retraining.

Option B is correct because configuring a knowledge base with Retrieval Augmented Generation (RAG) allows the chatbot to retrieve relevant documents from the company's internal knowledge base and use them as context for generating answers. This ensures the model's responses are grounded solely in the provided data, preventing reliance on its pre-training knowledge.

Historical bias occurs when training data reflects past societal biases, such as fewer women in certain job roles. This leads to the model learning those biases and producing unfair outcomes.

Lake Formation allows you to register S3 locations, create resource links, and grant permissions at column level. The other options are not specific to Lake Formation column-level access control.

Reinforcement learning involves an agent learning by interacting with an environment and receiving rewards/punishments, such as a robot learning to walk through trial and error.

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

When a knowledge base is attached to a foundation model on Amazon Bedrock, the model retrieves information from the data source to augment its responses. If the data source is not refreshed, the model will return outdated information even if the model itself is current. Option C directly addresses this by identifying the stale data source as the root cause.

Option D is correct because for image generation models like Stable Diffusion XL, the CFG (classifier-free guidance) scale controls adherence to the prompt, and a style prompt can enforce artistic consistency. Option A and B (temperature, top_p) are for text models. Option C (increasing steps) improves quality but not style consistency.

Amazon SageMaker Ground Truth provides built-in workflows for common tasks like image classification and object detection, which simplifies the setup of labeling jobs. These pre-built templates handle the UI and data formatting, allowing users to focus on the labeling instructions rather than building the labeling interface from scratch.

The context window includes both input and output tokens. With 7,500 input tokens and a desired 1,000 output tokens, the total would be 8,500, exceeding the 8,000 limit. The model cannot generate the full response.

Reducing chunk size can improve precision, and using embeddings with higher dimensions may capture more semantic nuance. Increasing chunk size typically reduces precision, and disabling chunking would break retrieval. Increasing max context length does not affect retrieval.

The core principles typically include fairness, robustness, explainability, privacy, transparency, and accountability. Profitability is a business objective, not an ethical principle of responsible AI.