Generative AI concepts: AIF-C01 study notes and practice questions
What to know about Generative AI concepts for the AIF-C01 exam, from the Fundamentals of Generative AI domain (24% of the exam guide), then 3 free questions to test it. Part of the AIF-C01 study guide.
This topic is the vocabulary of generative AI. Questions describe a behavior or a need and ask which concept explains it.
Key terms
Term
Meaning
Foundation model (FM)
A large model pre-trained on broad data that can be adapted to many tasks
Large language model (LLM)
A foundation model for text
Token
A chunk of text (a word or part of one) that models read and generate; usage is priced and limited in tokens
Context window
The maximum number of tokens a model can consider at once, including prompt and answer
Embedding
A list of numbers (a vector) that represents meaning; similar meanings have nearby vectors
Vector
The numeric form of an embedding, stored in a vector database for similarity search
Chunking
Splitting documents into smaller pieces before embedding them
Prompt
The input given to a model
Transformer
The neural network architecture behind LLMs; uses attention to weigh relationships between tokens
Diffusion model
A model that generates images by gradually removing noise
Multimodal model
A model that works with more than one type of input or output, such as text and images
Hallucination
A confident but false or invented answer
Context engineering
Designing what information the model receives (instructions, retrieved documents, history, tools) so its answers are good
Foundation model lifecycle
Data selection → model selection → pre-training → fine-tuning → evaluation → deployment → feedback. Most companies start at model selection, using a model someone else pre-trained.
Use cases
Summarization, chatbots and assistants, translation, code generation, image and video generation, search, recommendations and customer service agents.
Common traps
Embeddings vs tokens. Tokens are how text is split; embeddings are vectors that capture meaning, used for search.
Bigger context window = better. It allows more input, but costs more per request and can still miss details.
Diffusion vs transformer. Image generation from text typically uses diffusion; text generation uses transformers.
A recruitment site wants a search feature that finds CVs with a similar meaning to a job description, even when the candidate uses different words, such as 'pipeline developer' for 'data engineer'.
The index must cover two million CVs, so simple keyword matching is not enough.
Which concept makes meaning-based matching possible?
Choose one answer.
AA. Embeddings, which turn each text into a vector of numbers where similar meanings sit close together.
BB. Stop sequences, which tell the model where to end the text it generates.
CC. Temperature, which controls how random the model's generated text is.
DD. Tokens, which split each CV into word pieces that the model counts when it processes text for generation.
Show the answer and explanation
Answer: A
A. Right: Correct. An embedding model turns each CV and job description into a vector, and the site compares vectors to find close matches. Different words with a similar meaning produce vectors that sit near each other. Learn more: Amazon Bedrock Knowledge Bases (opens in a new tab)
B. Wrong: A stop sequence marks where generated output should end. It does not compare the meaning of two documents, so it cannot match CVs to a job.
C. Wrong: Temperature controls how varied generated text is. It has no role in comparing CVs with a job description.
D. Wrong: Tokens are pieces of text that models count and process. Splitting text into tokens does not capture meaning, so matching on them still misses different wording.
Explanation
Embeddings map text to numeric vectors where closeness reflects similarity of meaning, which lets a search find CVs that use different words for the same skills. Stop sequences, temperature and tokens control or measure generation and do not compare meaning. This question tests recognizing the concept behind semantic search.
A furniture retailer wants to create photo-realistic images of sofas in different room settings from short text descriptions, without photographing every combination.
Which type of generative model is BEST known for this kind of text-to-image generation?
Choose one answer.
AA. A regression model trained on past sofa prices to estimate what each sofa should cost.
BB. A clustering model that groups similar room photos into style families for easy browsing.
CC. A diffusion model that refines random noise into a new image, guided by the text prompt.
DD. A speech recognition model that converts spoken product descriptions into written text.
Show the answer and explanation
Answer: C
A. Wrong: Regression predicts a numeric value, such as a price. It does not create images from text descriptions.
B. Wrong: Clustering sorts existing photos into groups. It cannot produce a new image of a sofa that was never photographed.
D. Wrong: Speech recognition converts audio into text. It works in the opposite direction from generating an image from a description.
Explanation
Text-to-image generation is a generative task, and diffusion models are the standard approach for it. The retailer needs new images created from descriptions, not predictions or sorting of existing photos. This question tests recognizing which model type generates images.
An insurer's claims assistant gives vague answers.
The instructions are already clear, but the model never sees the customer's policy wording, the claim history, or the results of a claim-lookup tool. The team builds a layer that selects these items and assembles them into the model's input for each request, within a fixed token budget.
What is this practice called?
Choose one answer.
AA. Pre-training, because the model learns the policies from scratch on each live request.
BB. Hyperparameter tuning, because it adjusts the model's settings to the insurer's data.
CC. Prompt caching, because it stores earlier responses and reuses them for the same prompt prefix.
DD. Context engineering, because it decides which information the model sees in each request.
Show the answer and explanation
Answer: D
A. Wrong: Pre-training is the large initial training run on broad data. It happens once, not for each request, and it does not feed live policy text into a request.
B. Wrong: Hyperparameter tuning changes training settings to improve a model's fit. It does not choose which documents or tool results appear in each request.
C. Wrong: Caching speeds up and lowers the cost of repeated prompt prefixes. It does not decide which new information the model is shown.
D. Right: Correct. Context engineering selects and arranges the policy text, conversation history and tool results that fill the model's input. It works within the token budget the model allows. Learn more: Amazon Bedrock Knowledge Bases (opens in a new tab)
Explanation
A model answers from what it is shown in its input, so the quality of an answer depends on the context it receives. Context engineering designs what goes into that input, such as retrieved documents, history and tool results, within a token limit. Pre-training, tuning and caching address other parts of the system. This question tests recognizing the practice that shapes the model's input.
Want more on Generative AI concepts? The 5 full-length practice sets have 26 more Generative AI concepts questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →