SaveMyCert
Log in
5 of 5 free questions left today·for 10 a day
AIF-C01 · Domain 3

Applications of Foundation Models practice questions

Applications of Foundation Models is worth 28% of the AIF-C01 exam — the heaviest of the 5 domains. Design considerations for FM-based applications, prompt engineering, training and fine-tuning, and evaluating FM performance — the heaviest-weighted domain. 6 fully worked examples are further down this page, answers included.

Exam weight
28%
the heaviest of the 5 domains
Questions
116
across 4 topics
Free, no account
5/day
a free account allows 10 a day
Explanations
Every option
right and wrong

Build a practice session

5 free questions left today.

Domains

How many?

5 free questions left today.

Mode

Ready when you are

10 fresh questions drawn across 1 of 5 domains, in Learn mode.

Review what you missed

Every question you answer incorrectly, and every question you flag while practising, is saved here automatically. Finish a session and you can come back to re-drill just those.

6 sample Applications of Foundation Models questions, fully explained

Questions from the AIF-C01 bank mapped to domain 3, with the answer key and the reasoning behind every option. None of them repeat the examples on the main AIF-C01 practice page.

Question 1Applications of Foundation Models

A developer wants a foundation model in Amazon Bedrock to give answers grounded in company data, without writing code for chunking, embedding and retrieval. Which Amazon Bedrock capability provides this?

Choose one.

  • a
    Amazon Bedrock Prompt Management

    Prompt Management stores and versions prompts. It does not ingest data or retrieve passages for the model.

  • b
    Amazon Bedrock Knowledge Bases Correct

    Knowledge Bases is the managed RAG capability: it chunks and embeds a data source, stores the vectors, retrieves relevant passages and passes them to the model.

  • c
    Amazon Bedrock Guardrails

    Guardrails filters inputs and outputs and can check whether a response is grounded in supplied context, but it does not retrieve that context from company data.

  • d
    Amazon Bedrock Model Evaluation

    Model Evaluation compares and scores model outputs; it does not connect a model to a data source.

The concept

Amazon Bedrock Knowledge Bases is the managed way to build RAG on AWS.

Why that’s the answer

The need is RAG without building the pipeline yourself. Knowledge Bases handles ingestion, chunking, embedding, vector storage and retrieval, then supplies the retrieved passages to the model. Guardrails is the tempting near-miss because of the word 'grounded', but it checks or filters responses rather than fetching data; Prompt Management and Model Evaluation manage prompts and measure output quality.

How to reason it out
  1. Identify the need: answers from company data with no custom retrieval code.
  2. Map 'managed RAG' to Amazon Bedrock Knowledge Bases.
  3. Separate retrieval (Knowledge Bases) from safety checks (Guardrails) and prompt or evaluation tooling.

Exam tip: Managed RAG in Amazon Bedrock = Knowledge Bases; Guardrails checks answers, it does not fetch the data.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

Question 2Applications of Foundation Models

A manufacturer wants a foundation model to answer technicians' questions from 4,000 pages of product manuals that change every month. The manuals are far larger than the model's context window, and the company wants to avoid paying for any training job. Which approach fits?

Choose one.

  • a
    Fine-tuning the model on the manuals

    Fine-tuning is a paid training job and would have to be repeated each month as the manuals change.

  • b
    Retrieval Augmented Generation (RAG) Correct

    RAG retrieves only the passages relevant to each question, so the manuals never have to fit in one prompt, and it needs no training job; updated manuals are simply re-indexed.

  • c
    Continuous pre-training on the manuals

    Continuous pre-training is an even larger training job than fine-tuning, which the company wants to avoid.

  • d
    Including the full manuals in each prompt

    In-context learning needs no training, but 4,000 pages exceed the context window, and sending them with every request would be very costly.

The concept

When the source material is too large for the prompt and training is ruled out, RAG is the fitting customization approach.

Why that’s the answer

Two constraints decide it. 'Avoid any training job' removes fine-tuning and continuous pre-training. 'Far larger than the context window' removes placing all manuals in each prompt, the near-twin that also skips training. RAG passes the model just the few passages relevant to each question and picks up monthly changes by re-indexing, not retraining.

How to reason it out
  1. List the constraints: very large, changing source; no training cost.
  2. Eliminate the training-based approaches.
  3. Eliminate prompt stuffing because the content exceeds the context window.
  4. Choose RAG, which sends only relevant passages.

Exam tip: Too much content for the prompt and no training budget: retrieve the relevant passages with RAG.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

Question 3Applications of Foundation Models

A support assistant gives different wording, and sometimes different facts, each time the same question is asked. The team wants consistent, repeatable answers from the same model. Which inference setting change fits?

Choose one.

  • a
    Raise the temperature so the model samples from a wider range of words

    A higher temperature increases randomness, so answers would vary even more.

  • b
    Lower the temperature so the model favours the most likely next words Correct

    A lower temperature makes the model choose high-probability tokens, giving focused and repeatable output.

  • c
    Lower the maximum output tokens so each answer is shorter

    Maximum output tokens only caps how long an answer can be; the wording and facts inside a shorter answer still vary from run to run.

  • d
    Raise top-p so the model draws from a larger pool of candidate words

    A higher top-p widens the set of words the model may pick from, increasing variation rather than reducing it.

The concept

Temperature controls how random a model's word choices are.

Why that’s the answer

Inconsistent answers to the same question point to too much randomness in sampling. Lowering temperature concentrates choices on the most likely tokens, so output becomes focused and repeatable. Raising top-p is the near-twin that moves the wrong way (more candidates), and lowering maximum output tokens shortens answers without making them any more consistent.

How to reason it out
  1. Read the symptom: the same input gives varied output.
  2. Link variation to sampling randomness.
  3. Pick the change that reduces randomness: lower temperature.

Exam tip: Need consistent, factual answers: lower the temperature. Need variety: raise it.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

Question 4Applications of Foundation Models

A marketing team uses one foundation model for two jobs: extracting invoice fields, where it set a temperature of 0.1, and brainstorming many different campaign taglines from one prompt. It reused the same settings, and the tagline results are nearly identical on every run. Which change should the team make for the tagline requests?

Choose one.

  • a
    Lower the temperature to 0 for the tagline requests

    A temperature of 0 makes output even more deterministic, so the taglines would become more alike, not less.

  • b
    Switch the tagline requests to a model with a larger context window

    A larger context window lets a request carry more input; it does not add variety to the output.

  • c
    Raise the temperature for the tagline requests Correct

    The low temperature chosen for extraction is suppressing variety; raising it lets the model sample less likely words and produce more varied taglines.

  • d
    Raise the maximum output tokens for the tagline requests

    More output tokens allow longer responses but the wording would still be near-identical between runs.

The concept

Different tasks need different inference settings: extraction wants a low temperature, creative generation wants a higher one.

Why that’s the answer

The team reused a setting chosen for a deterministic task (field extraction) on a creative task. A low temperature makes the model pick the most probable words, so repeated runs look alike. Raising the temperature for the tagline requests restores variety. Lowering it further moves the wrong way, and a bigger context window or output limit changes how much text fits, not how varied it is.

How to reason it out
  1. Notice the tagline requests inherited a temperature of 0.1.
  2. Link near-identical runs to low randomness.
  3. Raise temperature only for the creative job.

Exam tip: Inference settings are per request: keep temperature low for extraction and raise it for brainstorming.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

Question 5Applications of Foundation Models

A startup wants to adapt a foundation model to a new task at the lowest cost. It has a handful of good worked examples and must avoid any up-front spend before the first request is served. Which approach fits?

Choose one.

  • a
    Fine-tuning on the labeled examples

    Fine-tuning is a training job, and a handful of examples is also too few to fine-tune well.

  • b
    Continuous pre-training on domain text

    Continuous pre-training is one of the most expensive approaches and needs large amounts of unlabeled data.

  • c
    In-context learning with the examples Correct

    Placing the examples in the prompt (few-shot, in-context learning) adapts behaviour with no training at all, the cheapest approach.

  • d
    RAG over a store of product documents

    RAG adds reference knowledge rather than teaching a task, and it needs documents to be embedded and indexed in a vector store first, which is an up-front cost.

The concept

In-context learning (prompt engineering) is the cheapest way to adapt a foundation model because it needs no training and no supporting infrastructure.

Why that’s the answer

The constraints are lowest cost and no up-front spend. Putting the worked examples in the prompt changes the model's behaviour for that request without touching its weights or building anything first. Fine-tuning would use the same examples but as a paid training job, continuous pre-training is far more expensive still, and RAG supplies reference documents (after an indexing step) rather than demonstrating how to do the task.

How to reason it out
  1. Spot the constraints: lowest cost, nothing to pay for before the first request.
  2. Recall the cost ladder from prompt engineering through RAG and fine-tuning to pre-training.
  3. Match the asset (worked examples of a task) to the rung that uses them directly: in-context learning.

Exam tip: Cheapest customization: put instructions and examples in the prompt; RAG and every training-based approach cost more.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

Question 6Applications of Foundation Models

A team's RAG application needs a managed AWS database that keeps document embeddings next to its existing relational tables and returns the closest vectors to a query embedding. Which should it use?

Choose one.

  • a
    AWS Glue Data Catalog

    The Glue Data Catalog stores table metadata for analytics; it does not store vectors or run similarity search.

  • b
    AWS Lake Formation

    Lake Formation governs access to a data lake in Amazon S3; it stores no vectors and runs no similarity search.

  • c
    Amazon Aurora PostgreSQL Correct

    Aurora PostgreSQL-Compatible Edition supports the pgvector extension, so it stores embeddings alongside relational tables and runs nearest-neighbour queries; it is one of the vector stores the exam names.

  • d
    Amazon Athena on Amazon S3

    Athena runs SQL queries over data in Amazon S3; it is not a vector database with nearest-neighbour search.

The concept

A vector database stores embeddings and returns the vectors closest to a query vector; Amazon Aurora PostgreSQL-Compatible Edition does this with the pgvector extension.

Why that’s the answer

The requirement is vector storage and similarity search in a relational database. Aurora PostgreSQL with pgvector stores embeddings in tables and runs nearest-neighbour queries, which is why the exam lists Amazon Aurora among the vector stores, alongside Amazon OpenSearch Service, Amazon Neptune (Neptune Analytics) and Amazon RDS for PostgreSQL. The other options are real data services, but they catalogue metadata, govern a data lake or query files in S3; none performs vector similarity search.

How to reason it out
  1. Recognise 'embeddings' and 'closest vectors' as a vector-search need.
  2. Note the extra requirement: alongside existing relational tables.
  3. Pick the relational service with vector support and rule out services that cannot search vectors.

Exam tip: Vector storage for RAG on AWS: OpenSearch Service, Aurora PostgreSQL or RDS for PostgreSQL (pgvector), or Neptune Analytics.

Foundation Model Design: Selection, RAG, and Customization (AIF-C01) — the lesson that teaches this.

What AIF-C01 domain 3 tests, topic by topic

The official exam guide breaks Applications of Foundation Models into 4 topics. The question bank follows the same split, so a weak topic shows up as a cluster of misses you can go back and read.

Published AIF-C01 practice questions per topic in Applications of Foundation Models
TopicWhat it coversQuestions
Describe design considerations for applications that use foundation models (FMs)Exam guide task 3.1 (AIF-C01). Selection criteria for FMs (cost, modality, latency, multi-lingual, model size, model complexity, customization, input/output length, prompt caching); the effect of inference parameters on model responses (temperature, input/output length); Retrieval Augmented Generation (RAG) and its business applications (Amazon Bedrock Knowledge Bases); AWS services for storing embeddings within vector databases (Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, Amazon RDS for PostgreSQL); cost tradeoffs of FM customization approaches (pre-training, fine-tuning, in-context learning, RAG, model distillation); the role of AI agents and their business applications.29
Choose effective prompt engineering techniquesExam guide task 3.2 (AIF-C01). Concepts and constructs of prompt engineering (context, instruction, negative prompts); prompt engineering techniques (chain-of-thought, zero-shot, single-shot, few-shot, prompt templates); benefits and best practices (response quality improvement, experimentation, guardrails, discovery, specificity and concision, using multiple comments); potential risks and limitations (exposure, poisoning, hijacking, jailbreaking); prompt versioning and management strategies with Amazon Bedrock Prompt Management.29
Describe the training and fine-tuning process for FMsExam guide task 3.3 (AIF-C01). Key elements of training an FM (pre-training, fine-tuning, continuous pre-training, distillation); methods for fine-tuning an FM (instruction tuning, adapting models for specific domains, transfer learning, continuous pre-training); preparing data to fine-tune an FM (data curation, governance, size, labeling, representativeness, reinforcement learning from human feedback [RLHF]).29
Describe methods to evaluate FM performanceExam guide task 3.4 (AIF-C01). Approaches to evaluate FM performance (human-in-the-loop evaluation, benchmark datasets, Amazon Bedrock Model Evaluation); relevant metrics (Recall-Oriented Understudy for Gisting Evaluation [ROUGE], Bilingual Evaluation Understudy [BLEU], BERTScore, LLM-as-a-judge); whether an FM effectively meets business objectives (productivity, user engagement, task engineering); evaluating the performance of applications built with FMs (RAG, agents, workflows); business objective alignment metrics for AI applications (task completion rate, user satisfaction, cost per interaction).29
Total116

Revise Applications of Foundation Models before you drill it

Other AIF-C01 domains

Applications of Foundation Models: your questions

Applications of Foundation Models is domain 3 of the AIF-C01 exam guide and carries 28% of the scored content — the heaviest of the 5 domains. On a 65-question paper that works out to roughly 18 questions, though AWS does not publish an exact per-domain count and individual exam forms vary.

Source

The domain weight and topic list on this page come from the official AIF-C01 exam guide.