Domain 2: Fundamentals of GenAI

AWS Certified AI Practitioner — Domain 2: Fundamentals of Generative AI

Exam Code: AIF-C01  |  Level: Foundational
Domain Weight: 24%  |  Total Domains: 5  |  Passing Score: 700/1000

---

Table of Contents

1. What is Generative AI?
   • 1.1 Traditional AI vs. Generative AI
   • 1.2 What Generative AI Can Produce
2. Foundation Models
   • 2.1 Definition and Key Characteristics
   • 2.2 Foundation Models vs. Traditional ML Models
   • 2.3 Models Available on Amazon Bedrock
3. Large Language Models
   • 3.1 How LLMs Are Pre-Trained
   • 3.2 LLM Capabilities
4. The Transformer Architecture
   • 4.1 Core Mechanism
   • 4.2 Encoder vs. Decoder vs. Encoder-Decoder
   • 4.3 Text Generation and Decoding Strategies
5. Tokens and Context Windows
   • 5.1 Tokenization
   • 5.2 Context Window
6. Embeddings and Vector Databases
   • 6.1 What are Embeddings?
   • 6.2 Vector Databases
7. Prompt Engineering
   • 7.1 Prompting Techniques
   • 7.2 Inference Parameters
8. Model Customization Techniques
   • 8.1 Technique Comparison
   • 8.2 RAG vs. Fine-Tuning Decision Guide
9. Types of Generative AI Models
10. Amazon Bedrock
    • 10.1 Core Features
    • 10.2 Bedrock Security and Pricing
11. Amazon Q
12. Hallucination and Output Quality
13. Exam Tips and Quick Reference

---

1. What is Generative AI?

Generative AI is a subset of artificial intelligence that creates new content — text, images, audio, video, code, or synthetic data — by learning patterns from vast training datasets. Unlike traditional ML, which classifies or predicts, generative AI produces novel outputs.

1.1 Traditional AI vs. Generative AI

1.2 What Generative AI Can Produce

---

2. Foundation Models

2.1 Definition and Key Characteristics

A Foundation Model (FM) is a large AI model trained on broad, internet-scale datasets using self-supervision, designed to be adapted across a wide range of downstream tasks. The term was introduced by Stanford HAI in 2021.

Emergent Abilities

Capabilities that appear in large models but are absent in smaller versions include multi-step reasoning, code generation, language translation, in-context (few-shot) learning, and chain-of-thought reasoning.

2.2 Foundation Models vs. Traditional ML Models

2.3 Models Available on Amazon Bedrock

Exam Tip: Amazon Titan models are AWS-native. Customer data submitted to Bedrock is never used to train the underlying foundation models — this is a critical security guarantee.

---

3. Large Language Models

Large Language Models (LLMs) are foundation models specifically pre-trained on massive text corpora to understand and generate human language.

• "Large" — billions to trillions of parameters
• "Language" — trained primarily on text
• "Model" — a statistical representation of language

3.1 How LLMs Are Pre-Trained

LLMs are trained using self-supervised learning on internet-scale text without manual labels.

RLHF — Reinforcement Learning from Human Feedback

1. Pre-train LLM on massive text corpus
         ↓
2. Supervised Fine-Tuning (SFT) on high-quality examples
         ↓
3. Train a Reward Model from human preference rankings
         ↓
4. Optimize LLM with RL (PPO) to maximize reward
         ↓
   Aligned, helpful, and safer model

3.2 LLM Capabilities

---

4. The Transformer Architecture

The Transformer, introduced in "Attention Is All You Need" (Vaswani et al., Google, 2017), is the architecture underlying all modern LLMs. Its key innovation is the self-attention mechanism, which allows every token to attend to every other token in parallel.

4.1 Core Mechanism

Self-Attention

For each token, self-attention computes how much weight to assign to every other token when producing its representation. This creates Query (Q), Key (K), and Value (V) vectors.

Attention(Q, K, V) = softmax(QKᵀ / √d) × V

Multi-Head Attention

Multiple attention heads run in parallel, each learning different types of relationships — syntactic, semantic, co-reference, and positional.

Positional Encoding

Because transformers process all tokens simultaneously (not sequentially), positional encodings are added to each token embedding to convey order information.

4.2 Encoder vs. Decoder vs. Encoder-Decoder

Key Concept: Most modern chat and code models use the decoder-only architecture. They generate text one token at a time, each new token conditioned on all previous tokens.

4.3 Text Generation and Decoding Strategies

LLMs generate text one token at a time, left to right. At each step the model outputs a probability distribution over the entire vocabulary.

---

5. Tokens and Context Windows

5.1 Tokenization

A token is the basic unit of text that an LLM processes. Tokenization splits raw text into tokens before the model sees it.

Common Tokenization Methods

Key Concept: Token count directly controls both context window usage and API cost. Input tokens and output tokens are each priced separately, with output tokens typically costing more.

5.2 Context Window

The context window is the maximum number of tokens — input plus output — the model can process in a single request.

---

6. Embeddings and Vector Databases

6.1 What are Embeddings?

Embeddings are dense numerical vectors that encode the semantic meaning of text (or other data). Semantically similar content produces vectors that are close together in the embedding space.

"king"   → [ 0.21, -0.40,  0.80,  0.11, ... ]
"queen"  → [ 0.20, -0.39,  0.79,  0.18, ... ]
"apple"  → [-0.32,  0.61, -0.10,  0.92, ... ]

Key Concept: Embeddings enable semantic search — finding documents by meaning rather than exact keyword match. They are the foundation of Retrieval Augmented Generation (RAG).

AWS Embedding Models (via Bedrock)

6.2 Vector Databases

Vector databases store embeddings and support fast Approximate Nearest Neighbor (ANN) search to find the most similar vectors to a query.

---

7. Prompt Engineering

Prompt engineering is the practice of designing and optimizing input text to get the best possible output from a generative AI model — without modifying model weights.

7.1 Prompting Techniques

Zero-Shot Prompting

Ask the model to perform a task with no examples provided.

Classify the sentiment of this review: "The shipping was late but the product is great."

Few-Shot Prompting

Provide labeled examples before the actual request to guide the model.

"I love this!" → Positive
"Terrible experience." → Negative
"It's okay." → Neutral
"This product changed my life!" →

Chain-of-Thought (CoT) Prompting

Instruct the model to reason step by step before answering. Dramatically improves performance on reasoning and math tasks.

Q: A store has 5 apples. 2 are sold and 3 more are added. How many remain?
A: Let me think step by step.
   Start: 5 apples
   After selling 2: 5 − 2 = 3 apples
   After adding 3: 3 + 3 = 6 apples
   Answer: 6

Other Key Techniques

7.2 Inference Parameters

Exam Tip: Temperature = 0 produces deterministic, consistent output — ideal for factual Q&A and classification. Temperature > 0 introduces randomness — ideal for creative writing and brainstorming.

---

8. Model Customization Techniques

When zero-shot and few-shot prompting are insufficient, the model itself can be customized. Techniques vary in cost, complexity, and whether they modify model weights.

8.1 Technique Comparison

Fine-Tuning Methods

8.2 RAG vs. Fine-Tuning Decision Guide

---

9. Types of Generative AI Models

Key Concept: Diffusion models are the dominant architecture for image generation. They work by learning to reverse a noise-addition process — starting from pure noise and iteratively denoising toward a coherent image conditioned on a text prompt.

---

10. Amazon Bedrock

Amazon Bedrock is a fully managed service that provides access to high-performing foundation models from multiple providers through a single, unified API — without requiring infrastructure management.

10.1 Core Features

Knowledge Bases (Managed RAG)

Bedrock Knowledge Bases provide a fully managed RAG pipeline.

Data Sources (S3, SharePoint, Confluence, web)
       ↓  Automatic chunking + embedding
Vector Store (OpenSearch Serverless, Aurora pgvector, Pinecone)
       ↓  At inference time
User Query → Embed → Retrieve top-K chunks → Inject into prompt → FM → Response

Agents

Bedrock Agents enable autonomous, multi-step task execution by orchestrating reasoning, knowledge base queries, and API calls via Lambda functions.

Guardrails

Bedrock Guardrails provide customizable safety controls applied at both input and output.

Model Evaluation

Additional Bedrock Features

10.2 Bedrock Security and Pricing

---

11. Amazon Q

Amazon Q is AWS's family of generative AI–powered assistants tailored to specific contexts.

Exam Tip: Amazon Q Developer replaced Amazon CodeWhisperer. It adds security scanning, code transformation (e.g., upgrade Java 8 → 17), and CLI completions beyond simple code suggestions.

---

12. Hallucination and Output Quality

Hallucination occurs when an LLM confidently generates plausible but factually incorrect information.

Hallucination Types

Mitigation Strategies

LLM Evaluation Metrics

---

Exam Tips & Quick Reference

Scenario-to-Answer Mapping

---

Common Traps

• RAG vs. Fine-Tuning: RAG is for grounding in current or proprietary data and requires no training. Fine-tuning is for teaching a specific style, tone, or vocabulary and requires training data and compute. The exam frequently conflates them.
• Temperature = 0 is not random: Temperature = 0 makes generation deterministic — the model always picks the highest-probability token. It does not mean the model refuses to answer.
• Tokens ≠ Words: Tokens are subword units. A single English word may be one or more tokens. API costs are billed per token, not per word.
• Amazon Q Developer ≠ Amazon Q Business: Q Developer is for writing code; Q Business is for enterprise knowledge management. They are different products with different pricing.
• Bedrock does not train on your data: Customer prompts and responses submitted to Bedrock are never used to train or improve AWS base models — a fundamental privacy guarantee.

---

Key Terms — Domain 2

---

End of Domain 2. Continue to Domain 3: Applications of Foundation Models →