Domain 3: Applications of Foundation Models

AWS Certified AI Practitioner — Domain 3: Applications of Foundation Models

Exam Code: AIF-C01  |  Level: Foundational
Domain Weight: 28%  |  Total Domains: 5  |  Passing Score: 700/1000

---

Table of Contents

1. Foundation Model Application Patterns
   • 1.1 Development Levels
   • 1.2 Application Selection Criteria
2. Retrieval Augmented Generation
   • 2.1 RAG Architecture
   • 2.2 Chunking Strategies
   • 2.3 Retrieval Methods
   • 2.4 Amazon Bedrock Knowledge Bases
3. AI Agents
   • 3.1 Agent Loop and Architecture Patterns
   • 3.2 Amazon Bedrock Agents
4. Text-Based Applications
   • 4.1 Summarization
   • 4.2 Classification and Information Extraction
5. Code Generation Applications
6. Image and Vision Applications
   • 6.1 Image Generation
   • 6.2 Image Understanding
7. Conversational AI and Chatbots
   • 7.1 Chatbot Types and Memory
8. Search and Information Retrieval
9. Document Processing and Analysis
10. Building GenAI Applications on AWS
    • 10.1 AWS Services for GenAI Applications
    • 10.2 AWS Architecture Patterns
11. Model Selection and Evaluation
    • 11.1 Choosing the Right Model
    • 11.2 Performance and Cost Optimization
12. Exam Tips and Quick Reference

---

1. Foundation Model Application Patterns

Foundation models enable a new application development paradigm — natural language replaces explicit logic, and a single model can power many different features.

1.1 Development Levels

The depth of customization and expertise required increases with each level.

1.2 Application Selection Criteria

---

2. Retrieval Augmented Generation

Retrieval Augmented Generation (RAG) solves two fundamental LLM limitations: knowledge cutoff dates and the absence of private or proprietary information. At inference time, relevant context is dynamically retrieved from an external knowledge base and injected into the prompt.

2.1 RAG Architecture

RAG operates in two distinct phases:

── INDEXING PHASE (runs offline) ─────────────────────────
  Documents (S3, SharePoint, web, etc.)
       ↓ Document loader
  [Text Splitter / Chunker]
       ↓
  [Embedding Model] → Dense Vectors
       ↓
  [Vector Database]  ← stored for retrieval

── RETRIEVAL PHASE (runs at inference time) ──────────────
  User Query
       ↓
  [Embedding Model] → Query Vector
       ↓
  [Vector DB: ANN Search] → Top-K Chunks
       ↓
  [Prompt Assembly] System + Retrieved Context + Query
       ↓
  [Foundation Model] → Grounded Response

Common RAG Quality Issues and Fixes

2.2 Chunking Strategies

How documents are split significantly affects retrieval quality. Smaller chunks produce more precise retrieval; larger chunks provide more surrounding context.

Note: A typical starting configuration is 512 tokens per chunk with 10–20% overlap between adjacent chunks to prevent missing context at boundaries.

2.3 Retrieval Methods

Advanced RAG Techniques

2.4 Amazon Bedrock Knowledge Bases

Bedrock Knowledge Bases provide a fully managed RAG pipeline — no infrastructure to provision or manage.

Supported Data Sources

Setup Flow

---

3. AI Agents

An AI agent is an LLM-powered system that autonomously understands a goal, plans a sequence of steps, calls external tools, observes results, and delivers a final response — often across multiple iterations.

3.1 Agent Loop and Architecture Patterns

ReAct Agent Loop

User Request
     ↓
[LLM: Reason — what step is needed?]
     ↓
[Identify tool or action]
     ↓
[Execute tool call]
     ↓
[Observe result] ──► More steps needed? ──► Yes ──► back to Reason
     ↓ No
[Final response to user]

Agent Architecture Patterns

Common Agent Tool Categories

3.2 Amazon Bedrock Agents

Bedrock Agents provide a fully managed agent runtime.

Key Concept: Bedrock Agents expose a full trace of every reasoning step, action, and observation — essential for debugging agent behavior and satisfying auditability requirements.

---

4. Text-Based Applications

4.1 Summarization

Abstractive summarization (LLM-generated) produces new prose that captures the meaning of the source. Extractive summarization pulls key sentences verbatim from the source.

Map-Reduce Pattern for Long Documents

Long Document
     ↓ Split into chunks
[Chunk 1] → [Summary 1]
[Chunk 2] → [Summary 2]   ─► [Final Summary]
[Chunk 3] → [Summary 3]

4.2 Classification and Information Extraction

Traditional ML vs. LLM for Classification

Common Information Extraction Tasks

---

5. Code Generation Applications

Amazon Q Developer is the primary AWS service for AI-assisted coding.

Supported Languages

Python, Java, JavaScript, TypeScript, C#, Go, Ruby, PHP, Rust, Kotlin, SQL, Shell, and more.

Exam Tip: Amazon Q Developer replaced Amazon CodeWhisperer. The security scanning and code transformation features are exclusive to Q Developer and not available in the original CodeWhisperer.

---

6. Image and Vision Applications

6.1 Image Generation

6.2 Image Understanding

Multi-modal LLMs such as Claude 3 Vision and Amazon Nova accept both text and image inputs.

Amazon Rekognition Key Features

---

7. Conversational AI and Chatbots

7.1 Chatbot Types and Memory

Chatbot Architecture Comparison

Amazon Lex vs. Amazon Bedrock for Chatbots

Conversation Memory Types

LLMs are stateless. The full conversation history must be passed with each request. Different memory strategies manage this at scale.

---

8. Search and Information Retrieval

Keyword Search vs. Semantic Search

Amazon Kendra

Amazon Kendra is a fully managed enterprise search service powered by ML.

Search Service Comparison

---

9. Document Processing and Analysis

End-to-End Document Pipeline

Raw Documents (PDFs, scans, Word files)
     ↓
[Amazon Textract] → Extracted text, key-value pairs, tables
     ↓
[Amazon Comprehend] → Entities, sentiment, key phrases, PII
     ↓
[Amazon Bedrock] → Summarization, Q&A, classification, generation
     ↓
Structured output stored in RDS, DynamoDB, or S3

Amazon Comprehend Feature Reference

Amazon Comprehend Medical

Specialized for healthcare and clinical text:

• Detects anatomy, medical conditions, medications, dosages, and test results
• Identifies Protected Health Information (PHI) — HIPAA-eligible service
• Maps findings to standard ontologies: ICD-10-CM, RxNorm, SNOMED CT

---

10. Building GenAI Applications on AWS

10.1 AWS Services for GenAI Applications

10.2 AWS Architecture Patterns

Pattern 1 — Simple Prompt Application

Best for content generation and basic Q&A with no private data.

User → API Gateway → Lambda → Amazon Bedrock → Response

Pattern 2 — RAG Application

Best for Q&A over private or proprietary documents.

Ingestion:  S3 → Bedrock Knowledge Base (chunk + embed) → OpenSearch

Retrieval:  User Query → Lambda → Bedrock RetrieveAndGenerate API → Response

Pattern 3 — Agentic Chatbot

Best for multi-turn conversations that require external actions.

User → API Gateway → Lambda
                         → Bedrock Agent
                               → Knowledge Base (retrieve context)
                               → Action Group → Lambda → External API
                         → DynamoDB (store chat history)
                    → Response with citations

Pattern 4 — Document Processing Pipeline

Best for automated extraction and analysis at scale.

S3 Upload Event
     → Lambda
           → Textract (extract text and structure)
           → Comprehend (entities, sentiment, PII)
           → Bedrock (summarize or classify)
           → DynamoDB / RDS (store results)
           → SNS (notify downstream systems)

Pattern 5 — Streaming Response

Best for conversational UIs where text appears progressively.

User → API Gateway (WebSocket) → Lambda → Bedrock (streaming) → Tokens streamed in real time

---

11. Model Selection and Evaluation

11.1 Choosing the Right Model

Decision Factors

Model Size vs. Quality Trade-off

Key Concept: Always right-size. Use the smallest model that meets your quality threshold. Overusing large models is the most common cause of unnecessary GenAI cost.

Bedrock Model Evaluation

11.2 Performance and Cost Optimization

Latency Optimization

Cost Optimization

---

Exam Tips & Quick Reference

Scenario-to-Answer Mapping

---

Common Traps

• Kendra vs. Bedrock Knowledge Bases: Kendra is enterprise search — it finds documents. Knowledge Bases power RAG — the FM reads retrieved chunks and generates an answer. These solve different problems.
• Bedrock Agents vs. simple LLM calls: If a scenario requires calling an external API, executing code, or taking a multi-step action, the answer is Bedrock Agents — not a plain Bedrock model invocation.
• Amazon Lex vs. Bedrock for chatbots: Lex handles structured intents and slot-filling (e.g., booking a flight). Bedrock handles open-ended reasoning. The keywords "intent," "slot," or "script" point to Lex.
• Textract vs. Comprehend: Textract extracts raw text and structure from documents (OCR). Comprehend analyzes text that is already extracted (NLP). Use them together in a pipeline.
• Model size and cost: Larger models are not always better. The exam may ask for the "most cost-effective" approach — the correct answer will use the smallest model that meets the stated requirements.

---

Key Terms — Domain 3

---

End of Domain 3. Continue to Domain 4: Guidelines for Responsible AI →