AWS Learning Material Certification Cheatsheets
AIF-C01 Domain 2 24% domain

Task 2.1: Explain the basic concepts of generative AI (GenAI)

Fundamentals of GenAI · 4,042 words · source: Certified-AI-Practitioner-AIF-C01/domain-2/task-2-1-cheat-sheet.md

Domain 2: Fundamentals of GenAI

Task Statement 2.1: Explain the basic concepts of generative AI (GenAI)

Domain 2 is 24% of scored exam content. Task 2.1 focuses on foundational GenAI vocabulary, model types, common use cases, the foundation model lifecycle, token-based cost thinking, context engineering, and basic agentic AI concepts.

Use this sheet to build the mental model for later Domain 2 and Domain 3 questions. Most exam questions for this task are matching, definition, ordering, or "which concept explains this scenario?"

Exam Objective Map

You should be able to:

AWS exam objective What to know for the exam
Define foundational GenAI concepts Tokens, tokenizers, context windows, chunking, embeddings, vectors, prompt engineering, LLMs, FMs, multimodal models, diffusion models
Identify GenAI use cases Text/image/video/audio generation, summarization, AI assistants, translation, code generation, customer service agents, search, recommendations
Describe the FM lifecycle Data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback
Explain token-based pricing Input and output tokens affect cost, quotas, latency, and context-window design
Describe context engineering Provide the model with the right instructions, examples, retrieved information, memory, tools, and constraints at inference time
Define foundational agentic AI concepts Agents, tools, memory, orchestration, workflow patterns, multi-agent systems, MCP, inter-agent communication

The Big Picture

Generative AI creates new content rather than only classifying, predicting, or detecting existing patterns.

AI
  Machine learning
    Deep learning
      Generative AI
        Foundation models
          Large language models
          Multimodal models
          Diffusion models
        Agentic AI systems can use FMs, tools, memory, APIs, and orchestration

Exam shortcut:

GenAI creates.
LLMs generate and understand language.
FMs are broad reusable models.
Embeddings turn meaning into vectors.
Context engineering decides what the model sees at inference time.
Agents use models plus tools to pursue goals.

Core GenAI Terms

Term Exam-ready definition Example
Generative AI AI that creates new content such as text, code, images, video, audio, or synthetic data. Draft product descriptions or generate images from text
Foundation model (FM) A large model trained on broad data that can be adapted to many downstream tasks. Use a text FM for summarization, Q&A, and classification
Large language model (LLM) A foundation model focused on language tasks, commonly based on transformer architecture. Chat assistant, document summarizer, code helper
Prompt The input instructions, content, question, examples, and constraints sent to a model. "Summarize this policy in five bullets."
Completion / response The output generated by the model from the prompt. Summary, answer, code, image description
Inference Running a model on new input to produce output. Invoke a Bedrock model with a user prompt
Token A unit of text used by a model; it can be a word, part of a word, punctuation, or other symbol depending on the tokenizer. "foundation" might be one token or split into subword tokens
Tokenizer Component that converts text into token IDs from the model's vocabulary. Convert a prompt into numeric input IDs
Vocabulary The set of tokens that a model can represent. Words, subwords, punctuation, special tokens
Context window The maximum amount of input and generated text a model can handle in one request or conversation turn. Prompt + retrieved context + chat history + output
Prompt engineering Designing prompts to improve model output quality. Add role, task, examples, format, and constraints
In-context learning Including examples or relevant information in the prompt so the model can adapt at inference time. Few-shot examples inside the prompt
Zero-shot prompting Ask the model to perform a task without examples. "Classify this review as positive or negative."
One-shot prompting Provide one example before the task. One labeled review, then a new review
Few-shot prompting Provide multiple examples before the task. Several input/output pairs in the prompt
Embedding A numerical vector representation of data that captures semantic meaning. Convert a paragraph into a vector for similarity search
Vector An ordered list of numbers representing features or meaning in a mathematical space. [0.041, 0.056, -0.018, ...]
Chunking Splitting larger content into smaller pieces so it can be embedded, retrieved, or fit into a context window. Split a PDF into paragraph-sized chunks
Vector database / vector index A store optimized for finding semantically similar vectors. Retrieve related policy chunks for RAG
RAG Retrieval Augmented Generation; retrieve relevant context and add it to the prompt before generation. Answer from company docs without retraining the model
Context engineering Designing the full model context: prompt, examples, retrieved data, memory, tools, constraints, and conversation state. Add user profile, policy snippets, system instructions, and tool definitions
Parameter A learned model weight adjusted during training. More parameters can increase capacity but also cost and resource needs. Billions of learned weights in an FM
Fine-tuning Supervised adaptation of a base model with task-specific examples. Teach a model to output a company-specific format
RLHF Reinforcement learning from human feedback; aligns model behavior with human preferences. Human-ranked responses improve assistant helpfulness

Tokens, Context, and Cost

Tokens matter because GenAI inference is commonly priced and limited by token usage.

Concept Why it matters for the exam
Input tokens The prompt, system instructions, examples, retrieved context, chat history, and tool definitions all consume tokens.
Output tokens Generated text also consumes tokens and can increase latency and cost.
Context window If the prompt plus context is too large, content must be shortened, chunked, summarized, or retrieved more selectively.
Token quotas Token usage can count against per-minute or per-day service quotas.
Model-specific tokenization Different models can count tokens differently, so token estimates are model-specific.
CountTokens API Amazon Bedrock can estimate input token usage before sending inference requests for supported models.

Cost and performance tradeoff:

More context can improve quality.
More context also increases token cost, latency, and risk of irrelevant information distracting the model.

Exam traps:

  • A longer prompt is not automatically better.
  • Input tokens and output tokens can both affect cost.
  • Token counting is model-specific.
  • Context-window size is a hard design constraint; use chunking, summarization, retrieval, or prompt compression when content is too large.

Embeddings, Vectors, and Chunking

GenAI models work with numbers, not raw human meaning. Embeddings and vectors are how text, images, or other data can be represented mathematically.

Concept What it does Exam signal phrase
Embedding model Converts text, images, or other data into vector embeddings. "Represent meaning numerically"
Vector space Mathematical space where similar meanings are closer together. "Semantic similarity"
Chunking Breaks large content into smaller pieces before embedding or retrieval. "Split documents into manageable pieces"
Vector index Stores vectors for similarity search. "Find the most relevant chunks"
Query embedding Converts the user's question into a vector for retrieval. "Compare query to document vectors"

Typical RAG-style flow:

Ingestion:
Documents -> chunks -> embeddings -> vector index

Runtime:
User question -> query embedding -> retrieve similar chunks -> augment prompt -> FM response

Exam shortcut:

Chunking prepares content.
Embeddings encode meaning.
Vectors enable similarity search.
Retrieved chunks provide context.

Transformer-Based LLMs

Transformers are the core architecture behind many modern LLMs.

Concept Exam-ready meaning
Transformer Neural network architecture designed to process sequences and capture contextual relationships.
Self-attention Mechanism that helps the model weigh which parts of the input are important when producing each output token.
Position embeddings / encodings Help the model understand token order.
Pre-training Large-scale training on broad data so the model learns language patterns and representations.
Fine-tuning Additional training to adapt a pre-trained model to a specific task, domain, style, or output pattern.

You do not need low-level math for the exam. Know that self-attention lets transformer models learn relationships between tokens across a sequence, which supports language understanding and generation.

Model Types and Architectures

Model or architecture What it is Best association
LLM Text-focused generative model, commonly transformer-based. Text generation, summarization, Q&A, translation, code
Foundation model Broad model that can support many tasks and be adapted. Reusable base model
Multimodal model Model that can process or generate more than one modality. Text + image, text + audio, image captioning, visual Q&A
Diffusion model Generative model that learns to reverse a noising process to create data. Image generation, upscaling, inpainting, some audio use cases
Stable Diffusion Diffusion approach that works in latent space instead of raw pixel space. Text-to-image generation
GAN Generative adversarial network with generator and discriminator networks. Older/common image generation architecture concept
VAE Variational autoencoder that learns compressed latent representations for generation. Latent-space generative modeling

Exam traps:

  • LLMs are not the only GenAI model type.
  • Multimodal means multiple data modalities, not just multiple languages.
  • Diffusion models are strongly associated with image generation, denoising, inpainting, and upscaling.
  • Transformers are strongly associated with modern LLMs and language generation.

Unimodal vs Multimodal

Type Meaning Example
Unimodal Uses one modality for input/output. Text-only LLM that takes text and returns text
Multimodal Uses multiple modalities as input and/or output. Image captioning, visual question answering, text-to-image generation

Common multimodal tasks:

  • Image captioning: image input -> text output
  • Visual question answering: image + question -> text answer
  • Text-to-image synthesis: text prompt -> image output
  • Image editing: image + instruction -> modified image
  • Speech or audio generation: text or prompt -> audio output

GenAI Use-Case Chooser

Use case What GenAI does AWS anchor
Text generation Writes, rewrites, expands, or transforms text. Amazon Bedrock
Summarization Compresses long content while preserving main ideas. Amazon Bedrock, Amazon Q
AI assistant Answers questions or helps complete tasks through conversation. Amazon Q, Amazon Bedrock
Translation Converts text between languages; GenAI can help with context-sensitive translation. Amazon Translate, Amazon Bedrock
Code generation Produces code snippets, functions, tests, or explanations. Amazon Q Developer
Customer service agent Handles customer questions, retrieves knowledge, and can trigger actions. Amazon Bedrock Agents, Amazon Lex, Amazon Q
Search Uses embeddings and retrieval for semantic search. Bedrock Knowledge Bases, vector databases
Recommendation engine Suggests items or content based on user/context signals. Amazon Personalize, SageMaker AI, Bedrock for natural-language recommendation explanations
Image generation Creates or edits images from prompts. Bedrock models, SageMaker JumpStart models
Video generation Generates or transforms video content. Multimodal GenAI model pattern
Audio generation Creates audio, voices, or music-like output. Multimodal/audio GenAI model pattern

Exam shortcut:

Generate content -> GenAI.
Answer from private documents -> RAG / knowledge base.
Use APIs and complete steps -> agent.
Need prebuilt deterministic translation or speech service -> managed AI service can be better.

Foundation Model Lifecycle

The exam guide names this lifecycle:

Data selection -> Model selection -> Pre-training -> Fine-tuning -> Evaluation -> Deployment -> Feedback
Stage Purpose What to remember
Data selection Choose relevant, high-quality, representative data. Data quality, bias, licensing, privacy, and harmful content matter.
Model selection Pick a base model or service that fits task, modality, cost, latency, context length, and risk. Do not default to the largest model.
Pre-training Train a model on broad data to learn general patterns. Expensive, compute-heavy, usually done by model providers or large organizations.
Fine-tuning Adapt a pre-trained model with task-specific examples. Useful when prompting/RAG is not enough for behavior, style, or format.
Evaluation Test quality, safety, alignment, and business fit. Use technical metrics, human review, benchmarks, and task-specific checks.
Deployment Integrate the model into an application, endpoint, API, assistant, or workflow. Consider latency, scale, monitoring, security, and cost.
Feedback Collect user signals, corrections, monitoring data, and human review. Drives prompt changes, retrieval changes, fine-tuning, or model replacement.

Practical project flow from Skill Builder:

Identify use case
  -> Experiment and select
  -> Adapt, align, and augment
  -> Evaluate
  -> Deploy and iterate
  -> Monitor

Exam traps:

  • Pre-training from scratch is usually the most expensive path.
  • Prompt engineering and in-context learning are often tried before fine-tuning.
  • RAG augments model input at inference time; fine-tuning changes model behavior through training.
  • Evaluation is iterative and should happen before and after deployment.

Adaptation Methods

Method Changes model weights? Best for Cost/complexity
Prompt engineering No Improve instructions, output format, and task clarity. Lowest
In-context learning No Teach by examples inside the prompt. Low, but uses tokens
RAG / knowledge base No Add private or current knowledge at inference time. Medium
Fine-tuning Yes Improve domain behavior, style, structured output, or specific task performance. Higher
Pre-training from scratch Yes Build a new general model from massive data. Highest

Memory hook:

Prompting changes instructions.
RAG changes context.
Fine-tuning changes behavior.
Pre-training creates the base capability.

Context Engineering

Prompt engineering is about the prompt text. Context engineering is broader: it is about everything the model receives and can use during inference.

Context element What it contributes
System instructions Role, rules, safety boundaries, tone, and task priorities
User prompt The user's immediate request
Conversation history Prior turns and continuity
Examples In-context learning and output patterns
Retrieved chunks Grounding from documents, databases, or knowledge bases
Memory Persisted facts or preferences across interactions
Tool definitions APIs/functions the model can call
Output constraints Format, schema, length, citations, or style
Guardrails Content, topic, safety, and policy controls

Good context engineering aims to:

  • Give the model enough relevant information to complete the task.
  • Keep irrelevant information out of the context window.
  • Balance quality against token cost and latency.
  • Ground responses in trusted sources when accuracy matters.
  • Make tool use and workflow steps explicit.

Exam trap: context engineering is not only "write a better prompt." It includes retrieval, memory, tools, orchestration, and constraints.

Agentic AI Concepts

An AI agent uses a model to reason over goals, context, and available tools, then chooses actions to complete a task.

Concept Exam-ready meaning Example
Agent AI system that can plan, choose actions, use tools, and pursue a goal. Support agent that checks order status and opens a ticket
Tool usage Agent calls an external function, API, database, workflow, or code interpreter. Call a CRM API
Workflow orchestration Coordination of steps, dependencies, retries, approvals, and handoffs. Route refund request to the right process
Memory management Store and retrieve relevant state across turns or tasks. Remember customer preferences or previous steps
Multi-agent system Multiple agents with specialized roles collaborate on complex tasks. Researcher, planner, coder, tester agents
Supervisor agent Agent that routes tasks to collaborator agents and combines results. Mortgage assistant routes to new-loan or existing-loan agent
Multi-agent communication Agents exchange messages, share memory, delegate tasks, or negotiate. Planner asks researcher for sources
MCP Model Context Protocol; a protocol pattern for connecting agents to external tools and systems. Bedrock AgentCore Gateway MCP configuration for external tools

Agentic Patterns

Pattern Best for Exam clue
Single agent with tools A focused task that needs external actions. "Call an API, retrieve info, update a record"
Workflow agent Predictable multi-step process with central orchestration. "Follow an approval workflow"
Supervisor + collaborators Complex task requiring specialized agents and routing. "One agent delegates to domain experts"
Peer multi-agent collaboration Open-ended reasoning where agents share, debate, or divide work. "Multiple autonomous agents negotiate or cooperate"

Exam traps:

  • A chatbot is not automatically an agent. It becomes agentic when it can plan, use tools, manage state, or orchestrate actions.
  • Multi-agent collaboration is useful for complex tasks, but it adds coordination, latency, cost, and governance complexity.
  • Tools need permissions, security controls, validation, and error handling.

AWS Service Anchors For Task 2.1

Task 2.1 is mostly conceptual, but know these service anchors:

Concept AWS anchor
Managed access to foundation models Amazon Bedrock
Experiment with and invoke FMs Amazon Bedrock playgrounds and runtime APIs
Embeddings and RAG with managed workflow Amazon Bedrock Knowledge Bases
Build agents that use tools and knowledge bases Amazon Bedrock Agents
Multi-agent collaboration Amazon Bedrock Agents multi-agent collaboration
MCP for external tools Amazon Bedrock AgentCore Gateway MCP configuration
Pretrained models and templates SageMaker JumpStart
Custom model building and deployment Amazon SageMaker AI
Code generation assistant Amazon Q Developer
Enterprise AI assistant Amazon Q Business

Do not over-focus on service implementation for Task 2.1. The exam objective is concept recognition.

Comparison Drills

Prompt Correct association
"Break a long document into smaller pieces before embedding." Chunking
"Represent paragraphs as lists of numbers for semantic search." Embeddings / vectors
"Find related documents by meaning, not exact keyword match." Vector similarity search
"Prompt plus retrieved chunks exceeds model limit." Context window / token limit problem
"Estimate request cost before inference." Token counting
"Include three examples in the prompt." Few-shot prompting / in-context learning
"Model handles text and images together." Multimodal model
"Generate an image by iteratively denoising random noise." Diffusion model
"Model decides to call an API to complete a task." Agent tool use
"Supervisor routes subtasks to specialist agents." Multi-agent collaboration
"Add current company policy text to the prompt at runtime." RAG / context engineering
"Train model on broad internet-scale data." Pre-training
"Adapt a base model with labeled company examples." Fine-tuning
"Improve model behavior with human preference feedback." RLHF

One-Page Memorization Version

  • GenAI creates new content: text, code, images, video, audio, or synthetic data.
  • Foundation models are broad reusable models; LLMs are language-focused FMs.
  • Tokens are the units models process; input and output tokens affect cost, quotas, latency, and context limits.
  • A tokenizer converts text into token IDs from the model vocabulary.
  • A context window limits how much prompt, retrieved context, chat history, and output fit in one request.
  • Embeddings are numerical vectors that capture semantic meaning.
  • Chunking splits large documents into pieces for embedding, retrieval, or context-window fit.
  • Vector similarity search finds semantically related chunks.
  • Transformers use self-attention and are the core architecture behind many modern LLMs.
  • Multimodal models use more than one modality, such as text plus images.
  • Diffusion models are strongly associated with image generation, denoising, upscaling, and inpainting.
  • FM lifecycle: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback.
  • Prompting changes instructions; RAG changes context; fine-tuning changes model behavior.
  • Context engineering includes prompts, retrieved data, memory, tool definitions, constraints, and orchestration.
  • Agents use models plus tools, memory, and workflow orchestration to pursue goals.
  • Multi-agent systems split complex work across specialized agents.
  • MCP helps connect agents to external tools and systems.
  • Bedrock is the AWS anchor for managed foundation models, RAG, and agents.

Mini Practice Questions

  1. A company wants to answer employee questions from internal policy documents without retraining a model. Which concept fits?

    • Answer: RAG / context engineering, commonly with Amazon Bedrock Knowledge Bases.

  2. A document is too large to fit into a model's context window. What technique helps prepare it for retrieval?

    • Answer: Chunking, followed by embeddings and storage in a vector index.

  3. A model converts a paragraph into [0.041, 0.056, -0.018, ...] so similar paragraphs can be found. What is this representation?

    • Answer: An embedding vector.

  4. A GenAI request includes a long system prompt, five examples, retrieved documents, chat history, and asks for a long answer. What cost factor increases?

    • Answer: Token usage. Both input and output tokens can affect cost and performance.

  5. A model can accept an image and answer natural-language questions about that image. What type of model is this?

    • Answer: Multimodal model.

  6. A model generates images by starting from noise and iteratively removing noise. What model class is this?

    • Answer: Diffusion model.

  7. A support assistant searches a knowledge base, calls an order-status API, and opens a case if needed. What makes it agentic?

    • Answer: It uses tools and orchestrates actions toward a goal.

  8. A team wants several specialized agents to work together, with one agent routing tasks to others. What pattern fits?

    • Answer: Multi-agent collaboration with a supervisor and collaborator agents.

  9. A company has tried prompt engineering and RAG, but the model still fails to follow a required output style. What adaptation method might help?

    • Answer: Fine-tuning with task-specific examples.

  10. What is the difference between prompt engineering and context engineering?

    • Answer: Prompt engineering designs the prompt text; context engineering designs the full inference context, including instructions, examples, retrieval, memory, tools, constraints, and workflow state.

Sources

Official AWS and course sources used: