Foundation Model Design: Selection, RAG, and Customization (AIF-C01)

Choosing a foundation model is the first design decision, and the exam expects you to recognize the criteria that guide it. No single model is best for every job, so you match the model to the need. The main selection criteria are summarized below.

The exam pattern is a scenario that stresses one criterion. If a question emphasizes fast responses for a chatbot, latency wins; if it stresses keeping spend low at high volume, cost and a smaller model win; if users write in several languages, multi-lingual support decides. Read for the constraint the scenario highlights and match the criterion.

Once you pick a model, inference parameters let you steer each response without changing the model itself. Inference is the act of generating an answer from a prompt, and these settings adjust how that answer is produced. The exam focuses on a few key levers.

The most tested lever is temperature. To make output more deterministic, consistent, and factual, such as for a support answer or data extraction, lower the temperature. To make output more creative and varied, such as for brainstorming marketing ideas, raise it. Top-p and top-k are alternative ways to control randomness and are usually adjusted at a concept level. Output length matters for both cost and completeness: too low a cap can cut answers off, while a high cap allows longer, more expensive responses. Remember that these settings change behavior per request; they do not retrain or permanently alter the model.

Retrieval Augmented Generation (RAG) is a technique that retrieves relevant information from an external knowledge source and adds it to the prompt, so the model answers from current or proprietary data instead of only what it learned during training. The flow is simple: the user asks a question, the system searches a knowledge source for the most relevant passages, those passages are inserted into the prompt as context, and the model generates an answer grounded in them.

RAG matters because a foundation model's built-in knowledge is frozen at training time and does not include your company's private documents. Without RAG, a model asked about your internal policy either cannot answer or may hallucinate, meaning it produces a confident but wrong answer. By grounding the response in retrieved facts, RAG reduces hallucination and keeps answers current, all without retraining the model.

Business applications are broad: internal knowledge assistants that answer from company wikis, customer-support bots grounded in product manuals, and research tools that cite source documents. On AWS, Amazon Bedrock Knowledge Bases is the managed RAG service. You point it at your documents, and it handles chunking, creating embeddings, storing them, retrieving the relevant pieces at query time, and passing them to the model, so you do not build the retrieval pipeline yourself. For the exam, connect the phrase "answer from current or private data without retraining" directly to RAG, and "managed RAG on AWS" to Amazon Bedrock Knowledge Bases.

RAG depends on finding passages by meaning, not exact keywords. To do that, each chunk of your documents is converted into an embedding, a numerical representation of meaning expressed as a vector, and stored in a vector database. When a user asks a question, the question is also embedded, and the database returns the stored chunks whose vectors are closest, meaning the passages most similar in meaning. Vector storage and similarity search are what make retrieval work.

The exam expects you to recognize which AWS services can store embeddings in vector databases. The named options are:

• Amazon OpenSearch Service — search and analytics engine with vector search capability, a common choice for RAG.
• Amazon Aurora — the managed relational database, which supports vector storage through PostgreSQL compatibility.
• Amazon Neptune — the managed graph database, which supports vector search.
• Amazon RDS for PostgreSQL — managed PostgreSQL, which stores and searches vectors using the pgvector extension.

You do not need to know how to configure these for the exam. The recognition-level point is that vector databases hold embeddings for similarity search, and that on AWS these four services can serve that role. When a question describes storing embeddings to power semantic search or RAG, any of these services is a valid answer, and Amazon Bedrock Knowledge Bases can use such a store behind the scenes as its managed vector database.

When a base model is not enough, you can customize it, and the approaches form a spectrum from cheap and easy to expensive and involved. The exam loves "which customization approach" questions, so learn the ladder.

Start at the top and move down only if you must. Prompt engineering and in-context learning are the cheapest way to customize because they require no training. RAG adds your data without retraining. Fine-tuning trains the model on your examples and costs more. Continuous pre-training and pre-training from scratch are the most expensive. Separately, model distillation creates a smaller, cheaper, faster model that mimics a larger one, trading a little quality for much lower inference cost. If a question asks for the cheapest way to adapt a model, the answer is prompt engineering or in-context learning; if it asks how to reduce ongoing inference cost, think distillation.

An AI agent is an application that uses a foundation model as its reasoning engine and can take actions to complete a goal, not just return text. An agent combines three ingredients: the FM for reasoning, tools it can call (such as looking up an order, querying a database, or triggering an API), and orchestration that decides the steps and chains them together. This lets an agent break a request into steps, call the right tool at each step, and use the results to reach an outcome.

Consider a customer-service example. A shopper asks an agent to change a delivery address. The agent interprets the request, calls a tool to look up the order, calls another tool to update the address, confirms the change, and replies in natural language. A plain chatbot could only describe how to do it; the agent actually does it by using tools and orchestrating the steps.

Business applications include automating customer-service tasks, coordinating multi-step workflows such as booking or IT requests, retrieving and combining information from several systems, and assisting employees with routine operations. For the exam, remember the defining idea: an agent is a foundation model plus tools plus orchestration that takes actions, which sets it apart from a model that only generates a response. The deeper design of multi-agent systems is a Domain 2 concept; here you need the role and the business value.

Task 3.1 questions usually describe a need and ask which design choice fits. A quick worked example ties the pieces together. Suppose a company wants an assistant that answers employee questions from its constantly changing internal HR documents, gives consistent factual replies, keeps cost low, and needs no data-science team.

Work through the decisions. Because the answers must come from current, private documents, the right technique is RAG, not fine-tuning, since RAG adds your data without retraining and stays up to date as documents change. On AWS, Amazon Bedrock Knowledge Bases provides this as a managed service, storing the document embeddings in a vector database such as Amazon OpenSearch Service. To keep replies consistent and factual rather than creative, set a low temperature. For model selection, favor a smaller, lower-cost model with adequate quality and low latency to control spend and keep responses fast. If the assistant also needs to take actions, such as submitting a leave request, add an agent with the right tools.

Notice the order of preference throughout: reach for the cheapest, least invasive option that meets the need. Prompt engineering and RAG come before fine-tuning; fine-tuning comes before continuous pre-training; and pre-training from scratch is almost never the answer for a business application. That instinct, match the constraint in the scenario to the lightest approach that satisfies it, is exactly what the exam rewards.