Fundamentals of GenAI practice questions

Chunking means dividing a large document into smaller pieces so an application can manage, store, and retrieve it.

A 200-page manual cannot fit in one context window, so splitting it into paragraph-sized sections is chunking, making option B correct. Tokenizing produces tiny model-level units, fine-tuning is a training step, and embedding converts content to numbers, so those distractors describe different operations.

1. Notice the action: a large document is broken into smaller sections.
2. Match that action to chunking.
3. Separate it from tokenizing, which operates at the level of small text units.

Exam tip: Chunking splits a large document into smaller pieces an application can handle.

An embedding is a numerical representation of a piece of content, created so similar meanings produce similar numbers.

Embeddings let applications compare meaning by comparing numbers, so option C is correct. A token is a text unit, the context window is a size limit, and pre-training is a training stage, so none of the other options describe an embedding.

1. Identify the key phrase: representing meaning as numbers.
2. Match it to the term embedding.
3. Eliminate token, context window, and pre-training as different concepts.

Exam tip: An embedding encodes meaning as numbers so similar meanings map to similar values.

A vector is the ordered list of numbers an embedding produces, and distance between vectors reflects similarity of meaning.

Embeddings place similar meanings at nearby points, so a short distance signals a strong semantic match, making option C correct. Semantic search does not rely on identical keywords, vector length is unrelated to closeness, and no re-training happens per query, so the distractors fail.

1. Recall that an embedding maps meaning to a vector.
2. Remember that similar meanings sit close together in that space.
3. Conclude that short distance means similar meaning, the basis of semantic search.

Exam tip: Similar meanings produce nearby vectors, so closeness in the vector space drives semantic search.

A diffusion model is a generative model used mainly for image generation.

Diffusion models start from random noise and gradually refine it into a clear image that matches the prompt, so option D is correct. LLMs generate text, and classification and regression models predict labels or numbers rather than creating images.

1. Identify the goal: generate images from text.
2. Match image generation to the diffusion model.
3. Rule out text and predictive models.

Exam tip: Diffusion models generate images, typically from a text prompt.

A multi-modal model handles more than one type of data, such as text plus images, in a single model.

Accepting an image and a text question together requires a model that spans both modalities, so option D is correct. A text-only model cannot read the image, a diffusion model is for image generation, and a tokenizer is not a reasoning model.

1. Note that the input mixes two data types: image and text.
2. Match multiple data types to a multi-modal model.
3. Eliminate text-only, image-generation, and tokenizing options.

Exam tip: A multi-modal model works with several kinds of data, such as text and images, at once.

A foundation model is a large, general-purpose model pre-trained on broad data and adaptable to many tasks.

The defining traits are breadth of training and reuse across tasks, so option B is correct. A single-task model, a rule-based system, and a vector database each miss the general, pre-trained nature of a foundation model.

1. Recall that a foundation model is trained broadly, not for one job.
2. Match broad, adaptable, pre-trained to option B.
3. Rule out narrow, rule-based, and storage descriptions.

Exam tip: A foundation model is broadly pre-trained and can be adapted to many tasks.