AIF-C01 Domain 2 - Fundamentals of Generative AI

Domain 2 is 24% of the scored AWS Certified AI Practitioner (AIF-C01) exam. It covers foundation-model vocabulary, lifecycle and economics, model selection, agentic AI, and AWS generative-AI options.

Core concepts

  • Token: a model’s unit of text input/output; it can be a character, subword, word, or punctuation sequence.
  • Context window: the maximum tokens a model can consider for a request, including instructions, conversation, retrieved context, tool results, and often the generated output.
  • Chunking: splitting source content into retrievable segments. Chunks that are too small lose context; chunks that are too large dilute relevance and consume the context window.
  • Embedding: a numeric vector representing semantic properties of data.
  • Vector: an ordered list of numbers. Similar vectors are found with distance/similarity measures for semantic search.
  • Transformer: a neural architecture based on attention that processes relationships among tokens and underpins most modern LLMs.
  • Multimodal model: accepts or generates more than one modality, such as text, image, audio, or video.
  • Diffusion model: learns to reverse a noise-adding process and is commonly used for image/audio/video generation.
  • Prompt engineering: designing the instructions and examples passed to a model.
  • Context engineering: managing the complete information environment around the model - system instructions, conversation, retrieval, memory, tool descriptions/results, and token budget.

Foundation-model lifecycle

  1. Select and govern broad pretraining data.
  2. Select an architecture and pretrain the base model.
  3. Adapt with instruction tuning, domain fine-tuning, continued pretraining, preference alignment, or distillation.
  4. Evaluate quality, safety, robustness, bias, cost, and business fit.
  5. Deploy for inference with the required latency, availability, and capacity.
  6. Collect feedback, monitor, and improve prompts, retrieval, guardrails, or model versions.

Most exam scenarios start from an existing FM. Training a foundation model from scratch is the most expensive and data-intensive option.

Token economics

  • Usage-based inference commonly charges separately for input and output tokens; long prompts, large retrieved contexts, and verbose answers raise cost and latency.
  • A larger context window is capability, not free memory. Include only relevant information.
  • Prompt caching can reduce repeated processing cost/latency for stable prompt prefixes when supported.
  • On-demand inference favors variable usage; provisioned throughput favors predictable capacity/latency but introduces commitment and utilization risk.
  • Smaller models are often cheaper/faster; larger or reasoning models may improve difficult-task quality.

Agentic AI

An AI agent uses a model to interpret a goal, plan, call tools, observe results, use memory, and continue until it returns an answer or completes an action.

  • Tool: a callable API, function, database query, browser, or code interpreter.
  • Memory: short- or long-term state that preserves relevant history beyond a single model request.
  • Workflow orchestration: controlling sequence, branching, retries, approvals, and failure handling.
  • Multi-agent system: specialized agents coordinate through supervisor, swarm, graph, or workflow patterns. It can improve separation of responsibilities but raises cost, latency, security, and observability complexity.
  • Model Context Protocol (MCP): an open protocol for exposing tools and contextual resources to compatible AI applications. MCP provides interoperability; authentication, authorization, and safe tool use still need explicit controls.

Use cases

Generative AI is suited to summarization, assistants, semantic search, content/image/audio/video generation, translation, code generation, recommendation explanations, document processing, and customer-service agents. It is less appropriate when output must be deterministic, perfectly factual, trivially produced by rules, or cannot be safely reviewed/validated.

Benefits and limitations

BenefitsLimitations and risks
adaptable across tasksplausible but false hallucinations
natural conversational interfacenondeterministic output
rapid prototyping and time to marketlimited explainability
creates and transforms unstructured contentbias, toxicity, privacy/IP risk
lowers the entry barrier to broad AI taskstoken cost, latency, and regional/model constraints

Grounding, validation, Guardrails, human review, and smaller task-specific models reduce different risks; no single technique guarantees accuracy.

Model selection criteria

  • modality and task capability;
  • quality on representative data, not marketing benchmarks alone;
  • context/input/output limits and supported languages;
  • latency, throughput, availability, and regional coverage;
  • token cost, customization cost, and operational complexity;
  • safety, licensing, data-use terms, explainability, and regulatory requirements;
  • model size/complexity and whether a traditional ML model is more appropriate.

AWS choices

  • Amazon Bedrock - managed API access to multiple FMs plus Knowledge Bases, Guardrails, evaluations, prompt management, and other GenAI building blocks.
  • Amazon SageMaker AI - deeper control over custom/open models, training, fine-tuning, hosting, and the ML lifecycle.
  • Amazon SageMaker JumpStart - pretrained models/algorithms deployable into SageMaker AI infrastructure.
  • Amazon Nova - Amazon’s multimodal FM family, accessed through Bedrock and customizable through supported AWS workflows.
  • Amazon Quick - AI workspace for research, analysis, content, applications, and automation.
  • Amazon Q - developer/AWS assistance; Q Business is closed to new customers after July 31, 2026, with Quick as the current enterprise-work successor.
  • Kiro - agentic coding environment.
  • Strands Agents - open-source SDK for tool-using and multi-agent applications.
  • Amazon Bedrock AgentCore - secure runtime and operational primitives for agents built with various models/frameworks.

Common traps

  • Embeddings represent meaning for retrieval; they do not contain the original document in human-readable form.
  • Chunking, embedding, retrieval, and generation are separate steps.
  • More context can increase cost and introduce irrelevant/conflicting evidence.
  • Agents add action-taking and orchestration around models; they are not simply chatbots.
  • Accessibility and speed do not remove governance, data-quality, or human-accountability requirements.

Sources