AIF-C01 · Domain 3 of 5

AIF-C01 Domain 3: Applications of Foundation Models

The heaviest AIF-C01 domain covers how foundation models are put to work: selection criteria, inference parameters, RAG and vector stores, customization costs, agents, prompt engineering, fine-tuning data and the methods and metrics used to evaluate FM applications.

Exam weight
28%
Official task statements
4
Estimated questions estimate
~14
Suggested study time estimate
16–22 hrs

About 12 min read Reviewed October 2026

Overview

Applications of Foundation Models has four task statements. Task 3.1 covers design decisions: FM selection criteria such as cost, modality, latency, size and prompt caching, inference parameters like temperature, Retrieval Augmented Generation with Amazon Bedrock Knowledge Bases, vector stores on OpenSearch Service, Aurora, Neptune or RDS for PostgreSQL, the cost of each customization approach, and the business role of AI agents. Task 3.2 covers prompt engineering techniques, best practices, risks such as jailbreaking and prompt versioning with Amazon Bedrock Prompt Management.

Task 3.3 covers training and fine-tuning: pre-training, continued pre-training, instruction tuning, distillation and preparing fine-tuning data, including RLHF. Task 3.4 covers evaluation with human review, benchmarks, Amazon Bedrock Model Evaluation, ROUGE, BLEU, BERTScore and LLM-as-a-judge, and the business metrics that show whether an FM application is worth running.

How it's tested

Domain 3 scenarios describe an FM-based application with a quality, cost or freshness problem and ask which design change fixes it. Options usually span prompting, RAG, fine-tuning and pre-training, or several inference parameters or metrics, and the correct choice is the lowest-effort option that fully meets the stated need.

  • Customization ladder

    A model lacks knowledge or style, and you choose prompt engineering, RAG, fine-tuning, distillation or pre-training based on cost, freshness and effort.

  • Tune the parameters

    Output is too random, too repetitive or cut off, and you identify the inference parameter, such as temperature or maximum length, to adjust.

  • Choose the evaluation

    A summarization, translation, RAG or agent application needs measuring, and you pick ROUGE, BLEU, BERTScore, LLM-as-a-judge, human review or a business metric.

Reasoning it through

An insurer wants a chatbot to answer from policy documents that change every month, and answers must cite current terms. Fine-tuning would freeze outdated wording into the model and require repeated training. RAG retrieves the current policy text at query time and passes it to the model as context. Amazon Bedrock Knowledge Bases manages ingestion, embeddings and retrieval, so monthly changes only require a data sync, not retraining.

Task statements 4

Official task statements from the Amazon Web Services outline, in outline order.

3.1 Describe design considerations for applications that use foundation models (FMs)FM selection, inference parameters, RAG, vector stores, customization costs and agents.
What it means
This task statement covers selecting an FM by cost, modality, latency, language support, size, complexity, customization options, input and output length and prompt caching. It also covers how inference parameters shape responses, what RAG is and where Amazon Bedrock Knowledge Bases fit, which AWS services store embeddings, the cost tradeoffs of pre-training, fine-tuning, in-context learning, RAG and distillation, and how AI agents serve business needs.
What to understand
Temperature controls randomness: low values give focused, repeatable answers and high values give varied, creative ones. Maximum output length caps the response and can cut answers short. RAG retrieves relevant chunks from a vector store and adds them to the prompt, keeping answers current without retraining. OpenSearch Service, Aurora PostgreSQL and RDS for PostgreSQL with pgvector, and Neptune can store vectors. Customization cost rises from in-context learning to RAG, fine-tuning and finally pre-training; distillation trains a smaller, cheaper model from a larger one. Prompt caching reuses repeated prompt prefixes to cut cost and latency.
How it shows up
A design need is described, and you choose the model, parameter, vector store or customization approach that meets it at the lowest cost.
Common confusion
Believing RAG changes model weights; it only adds retrieved context to the prompt at inference time.
3.2 Choose effective prompt engineering techniquesPrompt constructs, techniques, best practices, risks and prompt management.
What it means
This task statement covers the parts of a prompt (context, instruction, negative prompts), techniques such as zero-shot, single-shot, few-shot, chain-of-thought and prompt templates, the benefits and best practices of prompt engineering, risks including exposure, poisoning, hijacking and jailbreaking, and prompt versioning with Amazon Bedrock Prompt Management.
What to understand
Zero-shot gives only the instruction, single-shot adds one example and few-shot adds several, which helps the model copy a format. Chain-of-thought asks the model to reason step by step, improving multi-step answers. Negative prompts state what to avoid. Specific, concise prompts and iterative experimentation improve quality. Prompt injection and jailbreaking try to override instructions, and poisoning corrupts data the model relies on, so guardrails and input validation matter. Prompt Management stores, versions and shares templates so teams can test and roll back prompt changes.
How it shows up
A prompt produces poor or unsafe output, and you choose the technique, best practice or protection that improves it.
Common confusion
Confusing few-shot prompting with fine-tuning; few-shot puts examples in the prompt and changes no model weights.
3.3 Describe the training and fine-tuning process for FMsPre-training, fine-tuning methods, continued pre-training, distillation and data preparation.
What it means
This task statement covers the main elements of training an FM (pre-training, fine-tuning, continuous pre-training, distillation), fine-tuning methods such as instruction tuning, domain adaptation and transfer learning, and how to prepare fine-tuning data through curation, governance, sizing, labeling, representativeness and reinforcement learning from human feedback.
What to understand
Pre-training builds general ability from huge unlabeled datasets and is the most expensive step. Continued pre-training feeds more unlabeled domain text, such as legal or medical documents, to deepen domain knowledge. Fine-tuning uses smaller labeled prompt-response pairs to teach a task, format or tone; instruction tuning is a common form. Distillation trains a smaller student model to imitate a larger teacher, reducing inference cost. Fine-tuning data must be clean, representative and governed, and RLHF uses human preference ratings to align outputs with what people judge helpful and safe.
How it shows up
A customization goal and dataset are described, and you identify the training method or data preparation step that fits.
Common confusion
Mixing up continued pre-training and fine-tuning; continued pre-training uses unlabeled domain text, fine-tuning uses labeled examples.
3.4 Describe methods to evaluate FM performanceEvaluation approaches, FM metrics, business objectives and application-level evaluation.
What it means
This task statement covers ways to evaluate FMs, including human-in-the-loop review, benchmark datasets and Amazon Bedrock Model Evaluation, metrics such as ROUGE, BLEU, BERTScore and LLM-as-a-judge, judging whether an FM meets business objectives, evaluating applications built on FMs such as RAG systems, agents and workflows, and business alignment metrics like task completion rate, user satisfaction and cost per interaction.
What to understand
ROUGE measures word overlap with a reference and suits summarization. BLEU measures precision of n-grams against reference translations and suits translation. BERTScore compares meaning using embeddings, so it credits paraphrases. LLM-as-a-judge uses a model to grade outputs against criteria at scale, while human review catches nuance such as tone. Amazon Bedrock Model Evaluation runs automatic, judge-based and human evaluations and can assess knowledge bases. RAG systems need retrieval quality and answer faithfulness checked separately, and agents need task completion and tool-use accuracy measured.
How it shows up
An application and quality goal are described, and you select the evaluation method or metric that measures it.
Common confusion
Using BLEU for summarization; ROUGE is the standard summarization metric, while BLEU is mainly for translation.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • Temperature

    Inference setting controlling randomness of token selection.

    Remember
    Lower means more consistent output.
    Tested as
    Parameter questions.
    Trap
    Raising it to reduce hallucinations.

  • Retrieval Augmented Generation (RAG)

    Retrieves relevant data at query time and adds it to the prompt.

    Remember
    Keeps answers current without retraining.
    Tested as
    Design and cost questions.
    Trap
    Thinking RAG modifies the model.

  • Amazon Bedrock Knowledge Bases

    Fully managed RAG workflow that syncs documents, embeds them and retrieves relevant chunks.

    Remember
    Fastest path to RAG on AWS.
    Tested as
    Service selection questions.
    Trap
    Building a custom pipeline when managed RAG is asked for.

  • Vector store options

    OpenSearch Service, Aurora and RDS for PostgreSQL with pgvector, and Neptune.

    Remember
    All can hold embeddings.
    Tested as
    Storage questions.
    Trap
    Choosing DynamoDB or S3 Glacier for similarity search.

  • Customization cost order

    In-context learning, then RAG, then fine-tuning, then pre-training.

    Remember
    Pre-training is by far the costliest.
    Tested as
    Tradeoff questions.
    Trap
    Choosing pre-training for a style change.

  • Model distillation

    Training a smaller student model to mimic a larger teacher.

    Remember
    Cuts inference cost and latency.
    Tested as
    Cost questions.
    Trap
    Confusing it with fine-tuning for knowledge.

  • Few-shot and chain-of-thought

    Examples in the prompt; step-by-step reasoning instructions.

    Remember
    No weights change.
    Tested as
    Prompting questions.
    Trap
    Calling few-shot a form of training.

  • Prompt injection and jailbreaking

    Inputs crafted to override instructions or bypass safety rules.

    Remember
    Mitigate with guardrails and validation.
    Tested as
    Prompt risk questions.
    Trap
    Assuming a system prompt alone prevents it.

  • Amazon Bedrock Prompt Management

    Stores, versions and shares prompt templates.

    Remember
    Supports testing and rollback.
    Tested as
    Prompt governance questions.
    Trap
    Keeping prompts hard-coded in application code.

  • Continued pre-training vs. fine-tuning

    Unlabeled domain text versus labeled task examples.

    Remember
    Data type decides the method.
    Tested as
    Training questions.
    Trap
    Swapping the two.

  • ROUGE, BLEU, BERTScore

    Summarization overlap, translation precision, semantic similarity.

    Remember
    Match the metric to the task.
    Tested as
    Evaluation questions.
    Trap
    Using BLEU for summaries.

  • LLM-as-a-judge

    A model grades outputs against defined criteria.

    Remember
    Scales evaluation cheaply.
    Tested as
    Evaluation questions.
    Trap
    Assuming it fully replaces human review.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

Ways to customize an FM

ConceptMeaningKey differenceCommon trap
Prompt engineering / in-context learning Instructions and examples in the prompt Cheapest, no training Expecting it to add large knowledge
RAG Retrieve data at query time Fresh data, no weight changes Thinking it retrains the model
Fine-tuning Train on labeled task examples Changes weights for task or tone Using it for fast-changing facts
Continued pre-training Train on unlabeled domain text Deepens domain language Confusing it with fine-tuning
Pre-training Build a model from scratch Highest cost and data need Choosing it for minor needs

Prompting techniques

ConceptMeaningKey differenceCommon trap
Zero-shot Instruction only No examples Expecting strict formats
Few-shot Instruction plus examples Shows the pattern Calling it training
Chain-of-thought Ask for step-by-step reasoning Helps multi-step problems Using it for simple lookups
Prompt template Reusable prompt with variables Consistency across uses Not versioning templates

FM evaluation metrics

ConceptMeaningKey differenceCommon trap
ROUGE Overlap with reference text Summarization Using it for translation
BLEU N-gram precision vs. references Translation Using it for summaries
BERTScore Embedding-based semantic match Credits paraphrases Treating it as exact match
LLM-as-a-judge Model grades against criteria Scalable, flexible Skipping human spot checks

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Fine-tune for fresh data

    What candidates pick
    Fine-tune a model monthly on updated product documents.
    Why it looks right
    Fine-tuning sounds like teaching the model the content.
    Why it's wrong
    Fine-tuning is costly and goes stale between runs; RAG retrieves the latest documents at query time without retraining.

    Clue to watch for Source data changes often or answers must cite sources.

  2. Temperature fixes accuracy

    What candidates pick
    Raise temperature to get better factual answers.
    Why it looks right
    Tuning a parameter seems quick.
    Why it's wrong
    Higher temperature increases randomness; for factual consistency, lower temperature and ground the model with retrieved data.

    Clue to watch for Answers are inconsistent or invented.

  3. Pre-train for a tone change

    What candidates pick
    Pre-train a new model to match the company voice.
    Why it looks right
    A custom model seems the most thorough fix.
    Why it's wrong
    Pre-training needs vast data and compute; prompt instructions or light fine-tuning adjust tone at a fraction of the cost.

    Clue to watch for The need is style, format or tone.

  4. Few-shot is fine-tuning

    What candidates pick
    Label adding examples to a prompt as model training.
    Why it looks right
    Both use examples.
    Why it's wrong
    Few-shot examples live in the prompt and change no weights; fine-tuning updates the model with a training job.

    Clue to watch for The examples are placed in the prompt.

  5. BLEU for summaries

    What candidates pick
    Pick BLEU to score generated summaries.
    Why it looks right
    BLEU is a well-known text metric.
    Why it's wrong
    BLEU was designed for translation precision; ROUGE measures recall-oriented overlap suited to summarization.

    Clue to watch for The task is summarizing documents.

  6. One score for RAG

    What candidates pick
    Judge a RAG chatbot only on final answer quality.
    Why it looks right
    The answer is what users see.
    Why it's wrong
    Poor answers can come from weak retrieval or from generation, so retrieval relevance and answer faithfulness must be evaluated separately.

    Clue to watch for The application retrieves data before generating.

Answer framework

Apply these steps to Domain 3 questions about designing, customizing and evaluating foundation model applications.

  1. State the real gap: missing or changing knowledge, wrong format or tone, inconsistent output, high cost, or unsafe prompts, since each gap has a different fix.
  2. Try the cheapest fix first: clearer instructions, few-shot examples or chain-of-thought, and adjust temperature or length when output is random or truncated.
  3. Use RAG with Amazon Bedrock Knowledge Bases and a vector store when answers need private, current or citable data.
  4. Choose fine-tuning for task, format or tone with labeled examples, continued pre-training for domain language, and distillation to cut serving cost.
  5. Pick the evaluation: ROUGE for summaries, BLEU for translation, BERTScore or LLM-as-a-judge for meaning, separate retrieval checks for RAG and task completion for agents.

Practice question

A legal firm's assistant on Amazon Bedrock gives inconsistent answers to the same contract questions, and responses sometimes stop mid-sentence. The firm wants repeatable, complete answers without retraining. Which change should the team make?

Want more like this? The AIF-C01 practice test has questions for every domain.

Checklist

0 of 14

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Give this domain the biggest block of your schedule, since it is the largest share of the exam. Start with the customization ladder: write prompt engineering, RAG, fine-tuning, continued pre-training, distillation and pre-training in order of cost and effort, with the problem each one solves. Most Domain 3 questions become simple once you can place a scenario on that ladder.

Then get hands-on in Amazon Bedrock. Run the same prompt at low and high temperature, try zero-shot versus few-shot, and save a template in Prompt Management. If you can, create a small knowledge base from a few documents and ask questions that require them. Watching retrieval work explains RAG better than any diagram.

Next, learn the vector store options and the prompt risks, linking each risk to a mitigation such as guardrails or input validation. Finish with evaluation: build a table of ROUGE, BLEU, BERTScore, LLM-as-a-judge and human review, and add the separate checks needed for RAG and agents. Use practice questions to test whether you pick the cheapest option that fully solves the stated problem.

FAQ

How much of AIF-C01 is Domain 3?

Applications of Foundation Models is 28% of the scored content, the largest domain, with about fourteen of the fifty scored questions across four task statements.

When should I choose RAG over fine-tuning?

Choose RAG when answers must use private, frequently changing or citable data. Choose fine-tuning to change task behavior, format or tone using labeled examples.

What does temperature do?

It controls randomness in token selection. Low temperature gives consistent, focused answers; high temperature gives more varied and creative ones.

Which AWS services can store vector embeddings?

The guide names Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune and Amazon RDS for PostgreSQL. Bedrock Knowledge Bases can manage the store for you.

What is model distillation?

Training a smaller student model to reproduce the outputs of a larger teacher model, so inference becomes cheaper and faster with acceptable quality.

Which metric suits summarization?

ROUGE, which measures overlap with reference summaries. BLEU suits translation, and BERTScore measures semantic similarity.

How do you evaluate a RAG application?

Check retrieval relevance and coverage separately from whether the generated answer is faithful to the retrieved context, then track business metrics such as task completion.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 3 knowledge

You have covered FM selection, inference parameters, RAG, prompting, fine-tuning and evaluation. Test whether you can choose the lowest-cost fix that meets each scenario.

Take the AIF-C01 practice test Check your AIF-C01 readiness

Keep going