A. Incorrect.
Temperature controls how random or creative the model output is. It does not define how much input text the model can accept.

B. Correct.
The context window is the maximum amount of input and output, usually measured in tokens, that the model can process in one interaction.

C. Incorrect.
Batch size is related to processing multiple records together, usually during training or batch inference. It is not the prompt capacity limit.

D. Incorrect.
Model size describes the number of parameters in the model. It does not directly tell the team how much text fits in one prompt.

A. Incorrect.
Data poisoning affects the training data or model behavior before output is generated. It is usually an input or training-data attack.

B. Incorrect.
Parameter stealing targets the model itself by attempting to extract model weights or behavior.

C. Incorrect.
Model misuse is a broad governance risk, but it is not specifically an output vulnerability.

D. Correct.
Data leakage can occur when a model output exposes sensitive or confidential information that should not be returned to the user.

A. Incorrect.
Amazon Inspector helps identify software vulnerabilities. It does not provide private data-path connectivity between SageMaker and S3.

B. Incorrect.
Amazon Macie discovers sensitive data in S3. It does not manage network routing from notebooks to S3.

C. Correct.
A VPC configuration with an S3 endpoint keeps S3 access on the AWS network and avoids public internet paths.

D. Incorrect.
S3 Glacier Deep Archive is for long-term archival storage and is not appropriate for active notebook training data access.

A. Incorrect.
Fine-tuning can improve task-specific quality, but it requires labeled data and training cost. It is not the lowest-effort first step.

B. Incorrect.
Retraining a foundation model is expensive and usually outside the scope of an AI practitioner task.

C. Incorrect.
Training a new foundation model requires large datasets, specialized expertise, and significant compute cost.

D. Correct.
Prompt engineering can improve output quality by changing instructions and context without modifying the model.

A. Incorrect.
A custom Lambda loop adds unnecessary orchestration and can run into timeout, throttling, and retry complexity.

B. Correct.
Batch inference is designed for asynchronous processing of large prompt datasets when immediate responses are not required.

C. Incorrect.
Agents are useful for tool use and multi-step task execution, not simple bulk generation.

D. Incorrect.
Real-time inference is best when the application needs immediate responses, which is not required here.

A. Correct.
A negative prompt tells an image model which elements, styles, or objects to avoid in the generated image.

B. Incorrect.
Changing models might alter results, but it is not the direct control for excluding specific items.

C. Incorrect.
Higher temperature increases variability. It does not reliably prevent specific objects from appearing.

D. Incorrect.
A detailed prompt can guide desired content, but a negative prompt is the targeted method for exclusions.

A. Incorrect.
Using only the dominant demographic can make underrepresented groups perform worse in model predictions.

B. Incorrect.
Using only existing buyers can exclude new customers and reinforce historical behavior patterns.

C. Correct.
Balanced and diverse data helps the model learn from representative examples and reduces the chance of systematic bias.

D. Incorrect.
Public datasets can be useful, but they might not represent the retailer's users or fairness requirements.

A. Correct.
Fine-tuning uses labeled examples to adapt the model to the company's categories and expected tagging behavior.

B. Incorrect.
Prompt templates can help, but they do not fully use the labeled dataset to adjust the model.

C. Incorrect.
Zero-shot prompting uses no labeled examples and is usually less reliable for a custom taxonomy.

D. Incorrect.
Continued pre-training is generally based on unlabeled domain text and is not the best match for supervised category labels.

A. Incorrect.
Athena queries data in S3 with SQL. It is not a database service designed for vector similarity search.

B. Correct.
Aurora PostgreSQL can use pgvector-style vector support to store embeddings and run similarity queries for search applications.

C. Incorrect.
Redshift is a data warehouse. It is not the usual managed relational choice for embedding similarity search in this scenario.

D. Incorrect.
EMR is a big-data processing platform, not a database for online vector storage and lookup.

A. Correct.
Tokens are the pieces of text that a language model reads and generates.

B. Incorrect.
Vector representations are embeddings, not tokens.

C. Incorrect.
Model weights are learned parameters inside the model.

D. Incorrect.
Instructions sent to a model are prompts. Prompts are made up of tokens, but they are not tokens themselves.