Applications of Foundation Models
This is the largest domain on the exam. It is about putting foundation models to work: engineering effective prompts, customizing models through fine-tuning, grounding answers in your own data with retrieval augmented generation, and picking the right pre-built AWS AI service for a given task. Expect more than a quarter of scored questions here, many framed as practical scenarios.
Prompt engineering
Prompt engineering is the practice of designing and refining the instructions given to a model, its wording, context, examples, and structure, to obtain better, more reliable outputs, without changing the model's weights. Zero-shot prompting gives no examples; few-shot prompting includes a small number of worked input-output examples so the model can infer the desired pattern. Adding clear context and explicit constraints reduces ambiguity and helps the response match the user's intent and required format.
Customization: prompting vs. fine-tuning vs. RAG
There is a ladder of customization. Prompt engineering is fastest and cheapest and should be tried first. Fine-tuning updates a model's weights on your curated labeled data to reliably embed specialized knowledge, tone, or formats that prompting cannot achieve, at greater cost and effort. Retrieval Augmented Generation (RAG) grounds a model in your own documents by retrieving relevant passages at query time and adding them to the prompt, giving current, proprietary answers without retraining. Choosing among these based on cost, freshness, and depth of customization is a core exam skill.
Retrieval, embeddings, and vector databases
RAG relies on embeddings and a vector database. Documents are converted to embeddings and stored; at query time the system finds the passages whose embeddings are most similar to the question and injects them into the prompt. Amazon Bedrock Knowledge Bases manage this ingestion, embedding, and retrieval so a foundation model can cite current, organization-specific information. Amazon Bedrock Agents go a step further, letting a model plan and take multi-step actions by calling APIs and tools to complete a task.
Pre-built AWS AI services
AWS offers managed AI services that expose capabilities through an API with no model training required. For language and documents: Amazon Comprehend (entities, sentiment, key phrases in text), Amazon Textract (text, forms, and tables from scanned documents), Amazon Translate (language translation), Amazon Polly (text-to-speech), Amazon Transcribe (speech-to-text), Amazon Lex (conversational chatbots), and Amazon Kendra (intelligent enterprise search). For vision: Amazon Rekognition (objects, scenes, faces, and content moderation in images and video). Being able to map a described need to the correct service is heavily tested.
Trade-offs of managed services
The advantage of a managed AI service is production-ready capability without collecting data, training, or managing infrastructure; the trade-off is less control over the underlying model. For fully custom needs you would instead build on Amazon SageMaker AI. Managed services are also not infallible, they can make mistakes, so appropriate validation and human oversight still matter.