Fundamentals of Generative AI for the AWS AI Practitioner (AIF-C01): Foundation Models, Tokens, Embeddings & the GenAI Lifecycle

Generative AI is the reason the AWS Certified AI Practitioner (AIF-C01) exam exists in its current form. Domain 2 — Fundamentals of Generative AI — is worth roughly a quarter of your score, and it’s the domain most likely to trip up candidates who know traditional machine learning but haven’t stopped to pin down what a foundation model, a token, or an embedding actually is. The good news: this domain is conceptual, not mathematical. You won’t train a transformer on exam day — you need to explain, in plain terms, how generative AI works, what it’s good and bad at, and which AWS service you’d reach for.

This guide covers Domain 2 the way the exam tests it: definitions you can state confidently, the vocabulary that shows up in every question, the GenAI project lifecycle, and the AWS services that turn these concepts into products. If Domain 2 is your starting point, it pairs naturally with Domain 1: AI & Machine Learning fundamentals — generative AI is a subset of machine learning, so the two domains reinforce each other. For the full blueprint, see the AIF-C01 exam guide for 2026 and the domains breakdown.

What Generative AI Actually Is

Generative AI is a category of machine learning where models create new content — text, images, audio, code, video — rather than just classifying or predicting from existing data. A traditional ML model answers “is this email spam?” A generative model answers “write me a reply to this email.”

The distinction the exam wants you to internalize:

Traditional (discriminative) ML	Generative AI
Predicts a label or number	Produces new content
”Is this transaction fraud?"	"Draft a fraud-alert message”
Trained on labeled examples	Pre-trained on massive unlabeled data
Narrow, task-specific	Broad, adaptable to many tasks

Generative AI sits inside the broader ML family: AI ⊃ Machine Learning ⊃ Deep Learning ⊃ Generative AI. Every generative model is a deep-learning model, but not every deep-learning model is generative. That nesting is a favorite exam cue.

Foundation Models: the Engine of GenAI

The central concept in this domain is the foundation model (FM). A foundation model is a very large deep-learning model pre-trained on a huge, broad dataset so that it can be adapted to a wide range of downstream tasks. Instead of training a new model per task, you take one massive general-purpose model and steer it.

Key properties the exam expects you to know:

Pre-trained at scale. FMs learn general patterns of language, images, or code from enormous datasets, at a training cost only large organizations can afford.
Adaptable. A single FM can summarize, translate, answer questions, write code, and classify — all without task-specific training — because it learned general representations.
Expensive to train, cheap(er) to use. You almost never train an FM yourself. You consume one through an API or customize one with a small amount of your own data.

A Large Language Model (LLM) is a foundation model specialized for text — GPT-style, Claude, Amazon Titan Text, Llama, and so on. Multimodal models go further, accepting and/or producing more than one type of content (text and images, for example). On AWS, you access a catalog of foundation models from multiple providers through Amazon Bedrock — memorize that Bedrock is the managed, serverless way to call FMs.

How the Model Works: Transformers, Tokens & Prompts

You don’t need the math, but you do need the vocabulary. Here’s the chain the exam draws on.

Tokens

Models don’t read words — they read tokens. A token is a chunk of text: often a word, part of a word, or a punctuation mark. “Certification” might be one token or split into “Certif” + “ication” depending on the tokenizer. This matters for two exam-relevant reasons:

Pricing and limits are measured in tokens. Foundation-model APIs bill per input + output token, and each model has a maximum context window measured in tokens (how much text it can consider at once).
Cost control is a token conversation. “Reduce cost” often means “send fewer tokens” — shorter prompts, smaller context.

A useful rule of thumb: in English, 1 token ≈ 4 characters ≈ ¾ of a word, so ~100 tokens is roughly 75 words.

The Transformer and Next-Token Prediction

Modern foundation models are built on the transformer architecture, whose key innovation is the attention mechanism — it lets the model weigh how relevant every other token is to the one it’s generating. At its core, an LLM does something almost mundane: it predicts the next token over and over, given everything before it. Fluent essays, working code, and coherent answers all emerge from repeated next-token prediction. Understanding that generation is probabilistic explains the model’s biggest weakness (covered below).

Prompts and In-Context Learning

A prompt is the input you give the model. Because FMs are adaptable, you steer them largely through the prompt rather than through retraining. Two terms to know:

Zero-shot — you ask with no examples (“Classify this review as positive or negative”).
Few-shot — you include a handful of examples in the prompt to guide the format and behavior.

The craft of writing effective prompts is prompt engineering, and it’s important enough that we cover it separately in the AWS AI Practitioner prompt engineering guide. For Domain 2, know that prompting is the first and cheapest way to adapt a foundation model.

Embeddings and Vectors: How Models Understand Meaning

Embeddings are one of the most commonly missed Domain 2 concepts, so slow down here. An embedding is a numerical vector — a list of numbers — that represents the meaning of a piece of content (a word, sentence, image, or document) in a high-dimensional space. The crucial property: things with similar meaning have vectors that are close together.

"king"   -> [0.21, -0.63, 0.88, ... ]   (hundreds of numbers)
"queen"  -> [0.19, -0.55, 0.91, ... ]   close to "king"
"banana" -> [-0.72, 0.14, -0.30, ... ]  far away

Why the exam cares:

Semantic search. Convert a query and your documents to embeddings, then find the documents whose vectors are closest to the query’s — search by meaning, not keyword matching.
Vector databases. Embeddings are stored in vector databases so you can retrieve the most relevant chunks quickly. On AWS, vector storage options you should recognize include Amazon OpenSearch Service, Amazon Aurora / RDS for PostgreSQL with pgvector, Amazon Neptune, and Amazon DocumentDB.
RAG. Embeddings power Retrieval Augmented Generation — retrieving relevant context and feeding it to the model so it answers from your data. RAG is the standard fix for a model that doesn’t know your private or up-to-date information, and it appears throughout the exam.

If you take one thing from this section: embedding = meaning as numbers; similar meaning = nearby vectors.

The Generative AI Project Lifecycle

Domain 2 expects you to recognize the stages of building a generative AI application. It mirrors the traditional ML lifecycle but centers on adapting an existing FM rather than training from scratch:

Stage	What happens
1. Define use case	Identify the business problem and whether GenAI is the right tool
2. Select a foundation model	Choose an FM based on capability, modality, cost, latency, and context size
3. Adapt & customize	Improve results via prompt engineering → RAG → fine-tuning (in increasing cost/effort)
4. Evaluate	Measure quality, relevance, safety, and cost against your requirements
5. Deploy & integrate	Put the model behind an application or API
6. Monitor & improve	Watch quality, cost, latency, and drift; iterate

The order of adaptation techniques in stage 3 is a reliable exam point. Reach for them from cheapest to most expensive:

Prompt engineering — free, immediate, no data pipeline.
RAG — grounds the model in your own documents without retraining.
Fine-tuning — actually updates model weights on your labeled data; most powerful, most costly, requires ML effort.

When a question asks for the lowest-cost or fastest way to improve responses, the answer is usually prompt engineering or RAG — not fine-tuning or training a new model.

Real-World Use Cases

The exam rewards being able to match generative AI to appropriate business tasks. Common, exam-friendly use cases:

Text generation & summarization — drafting emails, summarizing documents, generating reports.
Chatbots & virtual assistants — customer support, internal help desks.
Code generation — writing, explaining, and translating code (e.g., Amazon Q Developer).
Image and media generation — marketing creatives, product imagery (Amazon Titan Image, Stable Diffusion via Bedrock).
Semantic search & Q&A over private data — RAG-based knowledge assistants.
Translation, personalization, and content moderation.

And, just as important, when not to use GenAI: for precise numerical calculation, deterministic rule-based logic, or tasks needing guaranteed-correct factual output, traditional software or classic ML is often the better, cheaper, safer choice.

Advantages and Limitations (Know Both Sides)

The exam is even-handed — it tests the benefits and the risks of generative AI. Be ready to argue both.

Advantages:

Adaptability — one model, many tasks.
Speed and scale — produces content far faster than humans.
Lower barrier to entry — consume an FM via API instead of building and training your own model.
Personalization — tailors output to context and user.

Limitations and risks:

Hallucinations — because the model predicts plausible tokens, it can produce confident, fluent, factually wrong output. This is the single most important limitation to name.
Nondeterminism — the same prompt can yield different answers (governed by parameters like temperature), which complicates testing.
Knowledge cutoff — an FM only “knows” its training data and has no awareness of recent or private information unless you supply it (that’s what RAG solves).
Bias and toxicity — models can reflect biases in their training data.
Cost and latency at scale — large models are compute-intensive.
Data privacy and IP concerns — what you send in prompts and how outputs are used matter.

Several of these bleed into other domains — bias, transparency, and fairness are the heart of Domain 4, covered in the responsible AI guide, and data protection sits in Domain 5, covered in security, compliance & governance. The exam deliberately connects them, so seeing the links helps.

The AWS Generative AI Stack

Finally, Domain 2 expects you to know where AWS services fit. Think of it as three layers:

Layer	Service	What it’s for
Infrastructure / build	Amazon SageMaker (+ JumpStart)	Build, train, fine-tune, and deploy your own models; JumpStart offers pre-built FMs and solutions
Foundation-model access	Amazon Bedrock	Serverless API access to FMs from Amazon and third parties; add RAG via Knowledge Bases and workflows via Agents
Applications	Amazon Q, PartyRock	Ready-made GenAI apps — Amazon Q for business/developer assistants; PartyRock for hands-on, no-code experimentation

A few service facts worth memorizing:

Amazon Bedrock is the default answer for “access foundation models without managing infrastructure.” It supports multiple model providers, Knowledge Bases for RAG, Agents for multi-step tasks, Guardrails for safety filtering, and model customization/fine-tuning.
Amazon SageMaker JumpStart is the answer when you want to deploy or fine-tune open foundation models yourself with more control.
Amazon Q is the family of AWS-built generative assistants (Q Developer for coding, Q Business for enterprise data).
PartyRock is an Amazon Bedrock playground for learning to build GenAI apps with no code — a good hands-on way to make these concepts concrete.

For a deeper tour of Bedrock specifically, the Amazon Bedrock guide for the AI Practitioner and the broader AWS AI services guide extend this section.

Turn Concepts into Confident Answers

Domain 2 is passable through understanding rather than memorization — but the exam phrases questions as short scenarios, so you need the vocabulary to fire instantly. “The model invented a fact” → hallucination. “Search by meaning” → embeddings / vector database. “Cheapest way to improve answers” → prompt engineering. “Answer from our private docs” → RAG. Building that reflex is exactly what practice questions are for.

Sailor.sh’s AWS AI Practitioner mock exams include generative-AI items covering foundation models, tokens and embeddings, the GenAI lifecycle, adaptation techniques, and the Bedrock/SageMaker/Amazon Q stack — each with an explanation of why the answer is right so you’re learning the concept, not just the key. Start with the free AIF-C01 practice questions to see where you stand, then build a plan with the AWS AI Practitioner study plan. Understanding generative AI well enough to explain it is what makes Domain 2 one of the most reliable scoring areas on the exam.

Frequently Asked Questions

What is Domain 2 of the AIF-C01 exam?

Domain 2 is Fundamentals of Generative AI, one of the five domains of the AWS Certified AI Practitioner exam and roughly a quarter of your score. It covers what generative AI is, foundation models, core concepts like tokens and embeddings, the generative AI project lifecycle, common use cases, the advantages and limitations of GenAI, and the AWS services (Bedrock, SageMaker, Amazon Q) that deliver it.

What is a foundation model?

A foundation model is a large deep-learning model pre-trained on a huge, broad dataset so it can be adapted to many downstream tasks — summarizing, translating, answering questions, generating code, and more — without task-specific training. Large Language Models (LLMs) are foundation models specialized for text. On AWS, you access foundation models through Amazon Bedrock.

What is the difference between a token and an embedding?

A token is a chunk of text (a word or part of a word) that a model reads and generates; usage and cost are measured in tokens. An embedding is a numerical vector that captures the meaning of content, positioned so that similar meanings are close together in vector space. Tokens are about how models process text; embeddings are about representing meaning for search and retrieval.

What are the main limitations of generative AI?

The big ones the exam tests are hallucinations (confident but factually wrong output), nondeterminism (the same prompt can give different answers), a knowledge cutoff (no awareness of private or recent data unless you provide it), potential bias and toxicity from training data, and cost/latency at scale. Retrieval Augmented Generation (RAG) addresses the knowledge-cutoff and private-data problems.

What’s the cheapest way to improve a foundation model’s responses?

Prompt engineering — refining the input you give the model — is the fastest and lowest-cost method, with no data pipeline required. If the model needs access to your private or current information, RAG is the next step. Fine-tuning, which updates the model’s weights on your labeled data, is the most powerful but also the most expensive and effort-intensive option. Reach for them in that order.

Which AWS service should I use to access foundation models?

Amazon Bedrock. It provides serverless, API-based access to foundation models from Amazon and third-party providers, plus Knowledge Bases for RAG, Agents for multi-step tasks, and Guardrails for safety. Use Amazon SageMaker (and JumpStart) when you want to build, fine-tune, or deploy models yourself with more control, and Amazon Q or PartyRock for ready-made and no-code generative AI applications.

Ready to lock in Domain 2? Drill realistic, explained questions with the Sailor.sh AWS AI Practitioner mock exams, then connect the dots with Domain 1: AI & ML fundamentals and the Amazon Bedrock guide.