Fundamentals of Generative AI — AWS Certified AI Practitioner AIF-C01 practice questions

Which BEST describes a foundation model?

1. A small classifier tuned to one specific task
2. A rules-based knowledge graph used for reasoning
3. A pretrained deep model adaptable to many tasks
4. A statistical technique for regression on tabular data
5. A cloud service that hosts container-based workloads

Answer: C — A pretrained deep model adaptable to many tasks

A) Foundation models are broad, not task-specific. B) Describes symbolic AI. C) Correct — trained on broad data, adapted (fine-tune/prompt/RAG) to many downstream tasks. D) Describes classical stats. E) Describes ECS/EKS/Fargate.

Which BEST describes prompt engineering?

1. Continuing pre-training on general web text
2. Updating the transformer's weight matrices offline
3. Increasing the model's context window at runtime
4. Redesigning the tokenizer to shrink vocabulary size
5. Designing inputs, examples, and constraints to guide outputs

Answer: E — Designing inputs, examples, and constraints to guide outputs

A) Continued pre-training. B) Fine-tuning. C) Not a technique — the window is architectural. D) Tokenizer change. E) Correct — role, context, examples, format, constraints.

What is the role of a vector database in a RAG system?

1. It stores relational rows for the application state
2. It caches raw HTTP requests before they hit the LLM
3. It indexes embeddings for fast similarity retrieval
4. It compresses model weights for smaller checkpoints

Answer: C — It indexes embeddings for fast similarity retrieval

A) That's an RDBMS. B) That's an HTTP cache. C) Correct — OpenSearch Serverless, Aurora pgvector, Pinecone, etc. D) Not a vector DB role.

Prompt engineering refers to:

1. Fine-tuning weights on a supervised dataset
2. Compressing model artifacts for edge deployment
3. Designing inputs to steer the model's outputs
4. Auto-scaling the endpoint based on token throughput

Answer: C — Designing inputs to steer the model's outputs

A) Fine-tuning. B) Model compression. C) Correct — role, context, examples, format, and constraints. D) Auto-scaling policy.

Approximately how many English words correspond to 1,000 tokens in common LLMs?

1. About 100 words per 1,000 tokens
2. About 250 words per 1,000 tokens
3. About 2,500 words per 1,000 tokens
4. About 750 words per 1,000 tokens

Answer: D — About 750 words per 1,000 tokens

A) Off by ~7x. B) Off by ~3x. C) Off by ~3x on the high side. D) Correct — ~0.75 word per token is the common rule of thumb.

Retrieval-Augmented Generation (RAG) reduces hallucinations by:

1. Lowering the model's temperature to near zero
2. Adding a prompt suffix that says 'do not hallucinate'
3. Retraining the model on a fresh general web crawl
4. Retrieving relevant documents and grounding the prompt

Answer: D — Retrieving relevant documents and grounding the prompt

A) Only reduces randomness. B) A wish, not a mechanism. C) Full retrain is neither RAG nor practical. D) Correct — grounded context makes the answer verifiable.

Bedrock available models?

Anthropic Claude, Meta Llama, Mistral, Cohere, AI21, Stability AI, Amazon Titan.

Chain of thought (CoT)?

Prompt model to show reasoning step-by-step. Improves complex reasoning.

Examples of Meta models on Bedrock?

Llama 3, Llama 3.1, Llama 3.2 (incl. multimodal).

RAG?

Retrieval-Augmented Generation. Retrieve relevant docs, include in prompt. Reduces hallucinations.