Applications of Foundation Models practice questions

Amazon Bedrock Knowledge Bases provides RAG as a managed service on AWS.

You point Amazon Bedrock Knowledge Bases at your documents and it handles chunking, creating embeddings, storing them, retrieving relevant pieces at query time, and passing them to the model, so option B is correct. Prompt Management (A) versions prompts, Polly (C) is text-to-speech, and Comprehend (D) does language analysis, none of which is managed RAG.

1. Identify the requirement as a managed RAG pipeline.
2. Recall that Amazon Bedrock Knowledge Bases fills that role on AWS.
3. Rule out services that do prompts, speech, or text analysis.

Exam tip: Amazon Bedrock Knowledge Bases is the managed RAG service on AWS.

To answer from private data without retraining, RAG is the low-cost design choice.

The requirement, answer from private documents, lowest cost, no retraining, describes RAG precisely: it adds the manuals to the prompt through retrieval without changing the model. Fine-tuning (A), continuous pre-training (C), and pre-training from scratch (D) all involve training and cost more.

1. Note the constraints: private data, lowest cost, no retraining.
2. Match 'answer from private data without retraining' to RAG.
3. Eliminate the training-based options as more expensive.

Exam tip: When a scenario needs answers from private data without retraining at low cost, choose RAG.

Temperature controls randomness; lower values yield more deterministic, factual output.

A consistent, factual, repeatable assistant needs low randomness, so a lower temperature is correct. A higher temperature (A) adds variety, max tokens (C) only limits length, and RAG (D) addresses grounding, not randomness.

1. Recall that lower temperature means more deterministic output.
2. Match 'consistent, factual, repeatable' to lower temperature.
3. Reject settings that add variety or address a different concern.

Exam tip: Lower the temperature for more deterministic, factual, repeatable responses.

Higher temperature increases randomness, producing more creative and varied output.

Brainstorming many varied taglines benefits from more randomness, so a higher temperature is correct. Lower or zero temperature (A, B) reduces variety, and temperature clearly does affect creativity, so D is wrong.

1. Recall the direction: higher temperature means more variety.
2. Match 'varied and creative' to a higher temperature.
3. Eliminate low, zero, and no-effect claims.

Exam tip: Raise the temperature when you want more creative, varied output.

The customization spectrum runs from prompt engineering (cheapest) to pre-training from scratch (most expensive).

Prompt engineering, also called in-context learning, adapts the model with no training and is therefore the cheapest, making A correct. Fine-tuning (A), continuous pre-training (B), and pre-training from scratch (D) all involve training and increasing cost.

1. Recall the ladder: prompt engineering, RAG, fine-tuning, continuous pre-training, pre-training.
2. Identify the no-training, lowest-cost end as prompt engineering.
3. Eliminate the training-based approaches.

Exam tip: Prompt engineering (in-context learning) is the cheapest customization and involves no training.

Vector databases store embeddings for similarity search; several AWS services can act as one.

Amazon OpenSearch Service supports vector search and stores embeddings for RAG, so C is correct. Polly (A), Transcribe (B), and Rekognition (D) are media and analysis services, not vector stores.

1. Recognize the need for a vector database that holds embeddings.
2. Recall the AWS vector-store options, including Amazon OpenSearch Service.
3. Eliminate speech and image services.

Exam tip: Amazon OpenSearch Service is one AWS service that stores embeddings for vector search in RAG.