Fundamentals of Generative AI practice questions

Worked examples for Fundamentals of Generative AI

1. Question 1

   A retail company stores thousands of product descriptions and wants a search feature that returns products that are **semantically similar** to a shopper's typed phrase, even when the shopper's wording does not match the product text word-for-word. The team plans to convert each product description into a numeric representation that captures its meaning, store those representations, and compare them by mathematical distance. What are these numeric representations called?

   1. A. EmbeddingsCorrect answer

      Correct. An embedding is a numerical vector representation of content whose position in vector space captures meaning, so semantically similar items sit close together and can be compared by mathematical distance — exactly what semantic search needs.

   2. B. Tokens

      Confuses tokens with embeddings. A token is a chunk of text (a word or subword piece) that the model reads and generates; token counts drive context limits and cost, but a token is not the numeric meaning representation used for similarity comparison.

   3. C. Prompts

      Confuses the input instruction with the stored representation. A prompt is the natural-language text a user supplies to guide a model's output; it is not a stored numeric encoding of a document's meaning.

   4. D. Hyperparameters

      Confuses training configuration with data representation. Hyperparameters are settings such as learning rate or batch size chosen before training; they describe how a model is trained, not the meaning of a piece of content.

   Explanation

   Embeddings are numerical vector representations of content — text, images, or other data — arranged so that items with similar meaning are close together in vector space, which is what makes distance-based semantic search possible. Tokens are the text chunks a model processes and are counted for context and cost, not compared for meaning. A prompt is the instruction text given to a model at inference time, and hyperparameters are training-time configuration settings; neither encodes the meaning of a stored document. See AWS — What is Generative AI? (https://aws.amazon.com/what-is/generative-ai/).

2. Question 2

   A media company wants a generative AI application that accepts a photograph together with a written question about that photograph, and returns a written answer. Which characteristic of a foundation model makes this possible?

   1. A. Tokenization — splitting the input into tokens lets the model convert images into text automatically.

      Confuses tokenization with multi-modality. Tokenization is how input is broken into discrete units the model processes; it is not what gives a model the ability to handle image data alongside text.

   2. B. Fine-tuning — a text-only model becomes able to read images once it is fine-tuned on labeled data.

      Misconception that fine-tuning adds new input modalities. Fine-tuning adapts a model's behavior on tasks within the modalities it already supports; it does not turn a text-only model into an image-understanding one.

   3. C. Multi-modality — the model can accept and work with more than one type of data, such as text and images, within the same model.Correct answer

      Correct. AWS describes generative AI foundation models as multimodal: a single model can be trained on and process multiple data types (text, images, audio, video), so it can take an image plus a text prompt and generate a text response.

   4. D. Diffusion — the diffusion architecture lets the model interpret both text prompts and image inputs to produce text answers.

      Confuses diffusion with multi-modality. Diffusion is a generative architecture that produces images by iteratively denoising; it is not the property that lets a model consume mixed input types and emit text.

   Explanation

   Foundation models can be multimodal, meaning one model is trained on and can process several kinds of data — text, images, audio, video — which is what allows a single model to take an image plus a text question and return a text answer. Tokenization only describes how input is chopped into units for processing, and diffusion is a specific image-generation architecture, so neither explains mixed-input capability. Fine-tuning adapts an existing model's task behavior rather than adding a data type the base model never supported.

3. Question 3

   A software team has shipped a document-drafting feature that calls a foundation model from one provider through Amazon Bedrock. The generative AI market moves quickly, and leadership wants to know how much rework would be needed if, six months from now, a model from a *different* provider turns out to perform better on the team's task. The application's business logic and hosting stay the same. Which statement accurately describes how Amazon Bedrock affects the effort of switching foundation models?

   1. A. Each model provider available in Amazon Bedrock requires its own SDK, endpoint, and commercial agreement, so moving to another provider's model means integrating a new client library and re-onboarding with that vendor.

      This is the misconception that Bedrock is a directory of vendor endpoints rather than an abstraction layer. Bedrock's defining characteristic is that models from multiple providers are reached through a single AWS API, with AWS handling the provider relationship and access — no per-vendor SDK or contract is required.

   2. B. Switching to a different provider's model requires re-training or fine-tuning that model on the team's data first, because a model cannot serve requests in Bedrock until it has been customized for the specific application.

      This confuses customization with basic access. Base foundation models in Bedrock are ready to invoke as-is; fine-tuning is an optional step that adds training and hosting cost, never a prerequisite for calling a model.

   3. C. Because Bedrock exposes foundation models from multiple providers through a single API, switching is largely a matter of changing the model identifier and adapting the request/response details and prompts for the new model, then re-testing quality — no new infrastructure has to be provisioned.Correct answer

      Correct. Bedrock's 'choice of models' advantage comes from unified, serverless API access to foundation models from several providers, so experimenting with and adopting a different model is a configuration-and-prompt-tuning exercise, not an infrastructure migration.

   4. D. Comparing and switching between foundation models is only possible in Amazon SageMaker AI; an application built on Bedrock is bound to the model chosen at design time and must be migrated to SageMaker to change models.

      This inverts the Bedrock/SageMaker ownership boundary. SageMaker is for building, training, and self-hosting models; Bedrock is precisely the service designed for choosing among and swapping between managed foundation models, and nothing binds a Bedrock application to one model permanently.

   Explanation

   Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models from several providers through a single, serverless API, which is why model selection stays a late and reversible decision: adopting a different model means pointing at a different model identifier, adjusting the request payload and prompts to that model's expectations, and re-evaluating output quality. It does not require a separate vendor SDK or contract, because AWS fronts the providers; it does not require fine-tuning first, because base models are directly invocable and customization is optional; and it does not require moving to Amazon SageMaker AI, which serves the different purpose of building, training, and self-hosting models. Prompts almost always need re-tuning per model, but the surrounding application and infrastructure do not change.

4. Question 4

   A logistics startup has no machine learning engineers. Its leadership wants to add a feature that drafts shipment-delay notification emails in several tones and languages. An advisor tells them that generative AI's advantages of simplicity, adaptability, and responsiveness make this feasible for their team. Which statement correctly explains one of those advantages in this scenario?

   1. A. Adaptability means a single foundation model automatically retrains itself on the startup's shipment data every time a user submits a prompt, so it continuously improves without any engineering effort.

      Confuses adaptability with continuous self-retraining. Adaptability means one general-purpose model can be applied to many different tasks; inference does not update the model's weights, and customization requires deliberate techniques such as fine-tuning.

   2. B. Responsiveness means the model returns an identical, reproducible email for a given prompt on every call, which is what allows it to be used in an automated workflow.

      Confuses responsiveness with determinism. Responsiveness refers to generating content quickly on demand; generative output is nondeterministic — the same prompt can yield differently worded results across calls when sampling is enabled.

   3. C. Because the startup has no ML engineers, the advantages only apply if it first labels a corpus of past emails so the foundation model can learn the required tones.

      Assumes labeled training data is a prerequisite. A foundation model is already pre-trained on broad data and can produce varied tones from instructions and examples supplied in the prompt, without a labeled corpus.

   4. D. Simplicity means the startup can guide a pre-trained foundation model with natural-language prompts instead of collecting a labeled dataset and training a task-specific model from scratch.Correct answer

      Correct. A key advantage of generative AI is that a pre-trained foundation model can be directed through natural-language instructions, removing the data-collection, labeling, and training work a bespoke model would require.

   Explanation

   Generative AI lowers the barrier to building AI features because a pre-trained foundation model is directed with natural-language instructions rather than a purpose-built, separately trained model — the simplicity advantage described in AWS's generative AI overview, alongside adaptability (one model serving many tasks) and responsiveness (fast, on-demand content generation). Adaptability does not mean the model retrains itself during inference, since generating a response leaves the model's weights unchanged. Responsiveness concerns speed of generation, not reproducibility — generative output is nondeterministic and the same prompt can produce differently worded results. And a labeled corpus is not a prerequisite, because the model's broad pre-training plus prompt instructions and examples are enough to vary tone and language.

5. Question 5

   A retail company deploys a customer-support assistant built on a large language model. During testing, reviewers notice the assistant occasionally states return-policy details that sound authoritative and well written but do not match any document in the company's policy library. Which statement BEST characterizes this behaviour and the realistic approach to managing it?

   1. A. This is a hallucination: the model generates statistically plausible text rather than retrieving verified facts, so grounding it in the policy library and adding human review reduce — but do not eliminate — the risk.Correct answer

      Correct. AWS describes generative AI as producing new content from patterns learned in training data, which can yield fluent output that is not factually grounded. Grounding techniques and human oversight lower the incidence of hallucination but cannot guarantee its absence, so residual risk must be managed.

   2. B. This is a data-retrieval defect: connecting the assistant to the policy library through retrieval augmentation will fully eliminate incorrect policy statements.

      Represents the misconception that retrieval augmentation eliminates hallucination. Grounding supplies relevant context and measurably reduces fabricated content, but the model still generates text probabilistically and can misread, over-generalize, or extend beyond the retrieved passages.

   3. C. This is a model-size problem: selecting a foundation model with substantially more parameters removes the possibility of fabricated answers.

      Represents the misconception that scale removes hallucination. Larger models are often more capable and can hallucinate less on some tasks, but fabrication is inherent to probabilistic generation, so no parameter count makes the output guaranteed factual.

   4. D. This is expected caching behaviour: the assistant is replaying stale responses from earlier conversations and clearing the session history will correct it.

      Confuses generation with response caching. The invented policy text is newly produced from learned patterns, not a stored prior answer being replayed, so clearing history addresses nothing about the underlying limitation.

   Explanation

   Generative AI creates new content by predicting likely continuations from patterns in its training data rather than by looking up verified records, so it can produce fluent, confident text that is factually wrong — the limitation known as hallucination. Grounding the model in authoritative source content and layering human review are the accepted mitigations, but because generation remains probabilistic they reduce frequency rather than remove the failure mode, so residual risk must still be managed. Framing the issue as a retrieval defect or a parameter-count shortfall assumes a technical fix that fully guarantees factual output, which no current technique provides. Treating it as replayed cached text misdescribes the mechanism entirely, since the erroneous policy wording is newly generated rather than stored.

6. Question 6

   A product manager is explaining why the company's document-search team stores **vectors** produced by an embedding model, and why the numbers in those vectors are not just arbitrary identifiers. Which statement correctly describes what a vector (embedding vector) is in the context of generative AI?

   1. A. A compressed copy of the original text that the model can decompress to recover the exact wording at any time.

      Confuses an embedding vector with lossless compression or with storing the raw text. A vector captures semantic characteristics, not a recoverable byte-for-byte copy of the source content.

   2. B. The number of tokens a foundation model is allowed to process in a single request.

      Confuses vector dimensionality with the token count / context length. Token limits govern how much text fits in a request; they say nothing about the numeric meaning representation.

   3. C. An ordered list of numbers that positions a piece of content in a mathematical space, so that content with similar meaning ends up numerically close together.Correct answer

      Correct. AWS describes embeddings as numerical (vector) representations of content in which semantically similar items are located near one another, which is what allows distance/similarity comparisons for search and recommendation.

   4. D. A unique random identifier assigned to each document so the database can look it up exactly.

      Treats the vector as an arbitrary primary key. Random IDs carry no meaning and cannot be compared for similarity, which is precisely the property embedding vectors provide.

   Explanation

   An embedding vector is an ordered list of numbers that encodes the meaning of a piece of content, placing it at a point in a high-dimensional space where semantically related items sit close together — that geometric closeness is what powers similarity search and recommendations. It is not a compressed or recoverable copy of the source text, it is not the request's token/context limit, and it is not an arbitrary identifier, because arbitrary identifiers cannot be compared for closeness of meaning (AWS — What is Generative AI?).

7. Question 7

   A data science team has built a novel transformer architecture in PyTorch and needs to train it from scratch on a large GPU cluster, control the training script and hyperparameters, and then deploy the resulting model to a real-time endpoint they manage. A separate application team in the same company only needs to send prompts to an existing third-party foundation model and receive completions, with no infrastructure to provision. Which statement correctly describes how Amazon Bedrock and Amazon SageMaker AI map to these two teams' needs?

   1. A. Amazon SageMaker AI suits the data science team because it is the platform for building, training, and hosting your own ML models with full control over the training environment, while Amazon Bedrock suits the application team because it provides serverless API access to foundation models with no infrastructure to manage.Correct answer

      Correctly states the ownership boundary: SageMaker AI gives you control over (and responsibility for) custom training and hosting, whereas Bedrock is a fully managed, serverless service that offers a choice of high-performing foundation models through a single API without provisioning infrastructure.

   2. B. Amazon Bedrock is the appropriate service for the data science team because Bedrock is designed for training custom model architectures from scratch on managed GPU clusters, while SageMaker AI is the API layer application teams use to invoke third-party foundation models.

      Reverses the ownership boundary between the two services. Bedrock does not let you bring your own model architecture and train it from scratch — it exposes existing foundation models (and customization such as fine-tuning on top of them) through an API — and SageMaker AI is the build/train/host platform, not the third-party FM invocation API.

   3. C. Both teams must use Amazon SageMaker AI, because SageMaker JumpStart is the only supported path for an application to send a prompt to a foundation model and receive a completion on AWS.

      Assumes SageMaker JumpStart is the sole route to foundation model inference. JumpStart is a hub for discovering and deploying models onto SageMaker infrastructure you manage; Amazon Bedrock independently offers serverless API access to foundation models, so it is not the only path.

   4. D. Both teams must use Amazon Bedrock, because any workload that involves a neural network on AWS — including training a proprietary architecture written in PyTorch — has to run through the Bedrock API.

      Treats Bedrock as a universal wrapper for all deep learning on AWS. Bedrock's scope is using and customizing available foundation models; training a proprietary architecture from a custom PyTorch script requires a build-and-train platform such as SageMaker AI.

   Explanation

   The dividing line is ownership of the model and the infrastructure. Amazon Bedrock is a fully managed service that makes high-performing foundation models from Amazon and leading AI companies available through a single API, serverless, so a team can send prompts and get completions without provisioning or managing any infrastructure. A team that authors its own model architecture and training script needs the build-train-host platform instead, where it controls the training environment and the endpoint it deploys to — that is Amazon SageMaker AI, not Bedrock. Claiming Bedrock trains custom architectures from scratch inverts this boundary; claiming SageMaker JumpStart is the only way to invoke a foundation model ignores Bedrock's API entirely; and claiming every neural-network workload must route through Bedrock overstates its scope, since Bedrock's role is to use and customize existing foundation models rather than to train arbitrary ones.

8. Question 8

   An insurance company has shortlisted four internal workloads and wants to implement exactly one of them with a foundation model. An architect reminds the team to pick the workload that plays to generative AI's strengths while staying clear of its known limitations. Which workload is the best fit for a foundation model?

   1. A. Calculating each customer's exact premium refund from policy and payment tables, where recalculating the same case must return a byte-identical figure every time.

      Misconception that a generative model is a deterministic calculation engine. With sampling enabled, the same prompt can yield different output across calls, so an exactly reproducible financial figure belongs in deterministic application or database logic, not in generation.

   2. B. Drafting first-pass claim summary letters from adjuster notes, which a claims handler reviews and edits before the letter is sent.Correct answer

      Correct. Producing new natural-language content from unstructured input is a core generative AI capability, and the human review step absorbs the residual risks of hallucination and output variability, so no limitation is violated (AWS, What is Generative AI?).

   3. C. Serving as the authoritative system of record for policy documents, replacing the document store, since the model memorized those documents during training and can be queried for them verbatim.

      Misconception that model weights are a queryable database. A foundation model encodes statistical patterns rather than retrievable records; asking it to reproduce authoritative documents invites fabricated text, which is why documents stay in a store and are supplied through retrieval.

   4. D. Generating the auditable, per-factor explanation of each automated claim denial that regulators can trace back through the decision.

      Misconception that a model's fluent self-explanation is a real audit trail. Large foundation models are hard to interpret, and the narrative they produce about their own reasoning is itself generated text, not a verifiable account of which factors drove the decision.

   Explanation

   Generative AI is strongest at creating new content — text, summaries, drafts — from unstructured input, and pairing that with human review keeps the residual risks of hallucination and variability contained. The limitations rule the other workloads out: nondeterministic sampling makes generation the wrong mechanism for a figure that must be exactly reproducible; model weights store learned patterns rather than retrievable authoritative records, so a model cannot act as a document system of record; and the interpretability limitation means a model-written rationale is generated prose, not a traceable audit trail of the factors behind a decision.