Module 1: Exploring Components of Generative AI Applications on AWS
Developing Generative AI Applications on AWS
3 Topics • ~45 minutes
📋 Module Agenda
Topic A: Generative AI Concepts
AI hierarchy, foundation models, pre-training, inference, context windows
Topic B: AWS Generative AI Stack
AI services, 3-layer stack, Amazon Bedrock, Amazon Nova
Topic C: Designing GenAI App Components
Core & optional components, architecture patterns, serverless example
🧠 GenAI: A Branch of Machine Learning
Key insight: GenAI models don't just classify or predict — they generate new content based on patterns learned from massive datasets.
🏗️ Foundation Models (FMs)
Pre-trained on broad, diverse datasets (internet-scale text, images, code)
Perform a wide variety of tasks without task-specific training
Can be fine-tuned for domain-specific use cases
Examples: Claude (Anthropic), Amazon Titan, Llama (Meta), Mistral
Range from small (Micro) to very large (Premier) sizes
Analogy: A foundation model is like a well-educated generalist — they've read extensively and can handle many subjects, but can specialize further with targeted study (fine-tuning).
⚙️ Pre-training & Inference
Key difference: Training builds the model's knowledge (expensive, one-time). Inference uses that knowledge to answer questions (fast, per-request).
📐 Context & Context Window
Context: The information available to the model in a single request-response exchange
Stateless: Each request is independent — the model doesn't "remember" previous conversations unless context is explicitly provided
Context Window: Maximum number of tokens (input + output) a model can process at once
Tokens: ~0.75 words per token (varies by model/tokenizer)
Model
Context Window
~Equivalent
Claude 3 Haiku
200K tokens
~150K words (a novel)
Amazon Nova Pro
300K tokens
~225K words
GPT-4o
128K tokens
~96K words
Analogy: The context window is like the size of your desk — it determines how many documents you can spread out and reference at the same time. A bigger desk (larger window) lets you consider more information simultaneously.
🔄 Foundation Model Functionality
Type
Input
Output
Examples
Understanding Models
Multimodal (text, images, docs)
Text only
Summarization, classification, Q&A, analysis
Creative Models
Multimodal (text, images)
Multimodal (text, images, audio, video)
Image generation, code writing, music, video
🧠 Understanding
Claude, Nova Micro/Lite — Analyze, extract, classify, reason about content
🎨 Creative
Nova Canvas, Stable Diffusion, Nova Reel — Generate images, videos, music
Lifecycle Change: Amazon Kendra is entering maintenance mode effective June 2026. New customers should use Amazon Bedrock Knowledge Bases for enterprise search and RAG use cases.
🏛️ AWS Generative AI Stack
Think of it as: Top layer = "use it", Middle layer = "build with it", Bottom layer = "build it yourself". Most developers work in the middle layer (Bedrock).
🚀 Amazon Bedrock — Key Benefits
🔌 API Access to 250+ FMs
Claude, Llama, Mistral, Amazon Nova, Stable Diffusion, Cohere, AI21 Labs — all via single API
🔒 Security & Privacy
Data stays in your account, encrypted, not used for model training. VPC endpoints, IAM integration
🎛️ Customizable
Fine-tuning, continued pre-training, RAG with Knowledge Bases, Guardrails for content control
🏗️ No Infrastructure
Fully managed — no servers, GPUs, or ML pipelines to manage. Pay per use (tokens/images)
Developer benefit: Focus on your application logic. Bedrock handles the heavy lifting of hosting, scaling, and serving foundation models.
✨ What's New: Amazon Nova Family⚠️ NEW
Launched December 2024 – First-party Amazon FMs optimized for Bedrock
Best balance accuracy/speed/cost. RAG, agents, code
Nova Premier
Multimodal → Text
Complex reasoning, long-context, hardest tasks
Nova Canvas
Text → Image
Image generation, editing, backgrounds
Nova Reel
Text → Video
Short video generation (6s clips)
Nova Sonic
Speech ↔ Speech
Voice-to-voice, real-time conversation
Why it matters: Amazon Nova provides cost-effective alternatives across the capability spectrum — up to 75% cheaper than comparable third-party models for many use cases.
🔗 What's New: OpenAI on Bedrock⚠️ UPDATE
Announced 2025 — GPT models now accessible through Amazon Bedrock
GPT Models on Bedrock
Access OpenAI's GPT-4o and other models using the same Bedrock Converse API and SDKs
Redesigned Console
New Bedrock console experience — model playground, comparison tools, unified navigation
Compatible APIs
Converse API works across all providers — same code for Claude, Nova, GPT, Llama, Mistral
Enterprise Governance
Use GPT with your existing AWS security controls — IAM, CloudTrail, VPC endpoints, Guardrails
Developer takeaway: Write once, switch models freely. The Converse API abstracts provider differences — your code works with any FM on Bedrock.
🧩 Designing GenAI Application Components
Core Components
Frontend / User Interface
Application Logic (orchestration)
Foundation Model (via Bedrock)
Prompt Template / System Prompt
Optional Components
Enterprise Data Sources (S3, DBs)
Embeddings Model
Vector Database (OpenSearch, Pinecone)
Memory / Conversation History
Tools & External APIs
Guardrails & Safety Filters
Design principle: Start with core components (prompt → FM → response). Add optional components incrementally as your application needs grow (RAG for enterprise data, agents for actions).
🏗️ Example: Serverless GenAI Architecture
Pattern: User → API Gateway → Lambda (orchestration) → Bedrock (inference) + Knowledge Bases (RAG) + DynamoDB (memory). Fully serverless, scales to zero.
🧪 Knowledge Check
Q1: Which layer of the AWS GenAI stack would a developer use to build a custom RAG application?
A) Applications Layer B) Tools & Services Layer C) Infrastructure Layer D) Networking Layer
B) Tools & Services Layer — Amazon Bedrock, Knowledge Bases, Agents, and Guardrails live here. This is where developers build custom GenAI apps without managing infrastructure.
Q2: What is the primary difference between "Understanding" and "Creative" foundation models?
A) Input format B) Output modality C) Context window size D) Cost
B) Output modality — Understanding models produce text-only output (analysis, classification). Creative models produce multimodal output (images, video, audio, code). Both can accept multimodal input.
📝 Module 1 Summary
GenAI Concepts
AI → ML → DL → GenAI hierarchy. FMs pre-trained on massive data. Context windows define working memory.
AWS GenAI Stack
3 layers: Apps, Tools/Services, Infrastructure. Bedrock = primary developer tool. 250+ FMs via single API.
Amazon Bedrock abstracts FM hosting. Focus on application logic, prompts, and data — not infrastructure.
⚠️ Lifecycle Changes & Updates⚠️ UPDATE
Important service changes since the original course materials were published
Amazon Kendra → Maintenance Mode (June 2026) Amazon Kendra will enter maintenance mode. No new features or customers accepted. Existing customers can continue use but should plan migration to Amazon Bedrock Knowledge Bases for enterprise search and RAG.
Amazon Q Business → Amazon Quick (Rebranding) Amazon Q Business has been rebranded to Amazon Quick. Same functionality, new name. All existing Q Business deployments continue unchanged.
✅ Recommended Actions
New RAG projects → Use Amazon Bedrock Knowledge Bases (not Kendra)
New enterprise assistant projects → Use Amazon Quick (formerly Q Business)
API integrations → Use the Converse API for model-agnostic code
First-party models → Evaluate Amazon Nova family for cost optimization