AIF-C01 GenAI Concepts, Tokens, Embeddings and Models Guide

This lesson gives you the GenAI vocabulary that appears all over AIF-C01. AWS wants you to know what tokens, embeddings, vector search, modalities, and foundation models are before it asks you to choose patterns like RAG or fine-tuning.

Token: Unit of text or other model input/output used in model processing and pricing.

Embedding: Numeric representation that captures semantic similarity so related items sit closer together in vector space.

Context window: The amount of input and output context a model can consider in one interaction.

High-yield GenAI map

Why these concepts matter together

These are not isolated glossary words. AIF-C01 uses them as a chain:

• tokens affect context size, latency, and cost
• embeddings support similarity search and retrieval
• multimodal models change what kind of input or output is possible
• foundation models are the broad starting point, not automatically the final business design

Retrieval vocabulary that shows up later

Common traps

• confusing embeddings with the original raw text itself
• assuming token count matters only for model quality and not for cost or context limits
• treating every model as text-only
• forgetting that foundation models are broad starting points, not always final business solutions on their own

Harder scenario question

A team wants users to search thousands of internal documents by meaning rather than by exact keyword matches. Which concept is the strongest foundation for that design?

• A. Embeddings
• B. Route tables
• C. Reserved capacity
• D. Basic IAM only

Correct answer: A. Semantic retrieval depends on turning content into embeddings that can be searched by similarity.

Decision order that usually wins

1. Decide whether the stem is about tokens, model context, modality, or FM capability.
2. Read token and context-window clues before blaming prompt style alone.
3. Match the task to model modality and capability rather than to hype.
4. Keep model access and prompt content separate from underlying context limits.
5. Choose the model lane that fits the output form and context requirement.

Quiz