Amazon Bedrock Complete Study Guide

AWS Certified AI Practitioner Exam Preparation

---

Table of Contents

1. Inference & APIs
2. Knowledge Bases (RAG)
3. Agents
4. Guardrails
5. Model Customization
6. Flows & Prompt Management

---

1. Inference & APIs

Core API Comparison

InvokeModel vs Converse API

Aspect	InvokeModel	Converse API
Purpose	Direct, low-level model invocation	Unified conversational interface
Format	Model-specific request/response	Model-agnostic format
Best For	Embeddings, images, model-specific features	Chatbots, multi-turn, tool use
Streaming	InvokeModelWithResponseStream	ConverseStream

Key Insight: Converse API automatically applies conversation templates for Mistral AI and Meta models.

Limitation: Converse API does NOT support embedding or image generation models.

API Selection Guide

Use Case	Recommended API
Multi-turn conversations	Converse
Tool use / function calling	Converse
Model-agnostic code	Converse
Embeddings generation	InvokeModel
Image generation	InvokeModel

---

Inference Parameters

Randomness Controls

• Temperature: Lower = deterministic, Higher = creative
• Top K: Limits to K most probable tokens
• Top P (Nucleus): Limits to tokens comprising top P% probability mass

Note: Default values vary significantly by model - always check model docs.

---

Tool Use (Function Calling)

Critical: The model ONLY requests tool calls - it NEVER executes them. Your application handles all execution.

User Query → Model → Tool Request → Your Code Executes → Tool Result → Model → Final Response

---

Prompt Caching

Mechanics

• TTL: 5 minutes (resets on each hit)
• Minimum tokens: 1,024 - 4,096 (model-dependent)
• Maximum checkpoints: 4 per request

Model Requirements

Model	Min Tokens	Cacheable Fields
Claude Opus 4.5	4,096	system, messages, tools
Claude 3.7 Sonnet	1,024	system, messages, tools
Amazon Nova	1,000	system, messages (text only)

Best Practice: Place cache checkpoints at END of static content.

---

Batch Inference

• Use CreateModelInvocationJob for large-scale processing
• Not supported with provisioned throughput
• More cost-efficient than individual API calls

---

Inference Profiles

Types

Type	Description
System-defined (Cross-Region)	Pre-configured multi-region routing
Application (Single-Region)	User-created for tracking
Application (Multi-Region)	User-created using system profile

Pricing: Based on calling region, NOT destination region.

---

Cross-Region Inference

Mode	Data Boundaries	Use Case
Geographic	US, EU, APAC only	Compliance/data residency
Global	Any region	Max throughput, ~10% savings

Security: No public internet traversal - stays on AWS internal network.

Important: Can route to regions NOT manually enabled in your account.

---

Service Tiers

Tier	Performance	Cost	Use Case
Reserved	Guaranteed	Fixed premium	Mission-critical
Priority	Fastest on-demand	Higher	Customer-facing
Standard	Baseline	Standard	General tasks
Flex	May queue	Lowest	Batch processing

Key: On-demand quota is SHARED across priority, standard, and flex tiers.

---

Latency-Optimized Inference (Preview)

Supported Models (limited):

• Amazon Nova Pro (us-east-1, us-east-2)
• Claude 3.5 Haiku (us-east-2, us-west-2)
• Llama 3.1 405B/70B (us-east-2, us-west-2)

Tradeoff: Falls back to standard inference when quota reached.

---

2. Knowledge Bases (RAG)

How RAG Works in Bedrock

Two Phases

Phase 1: Pre-processing (Ingestion)

Raw Data → Parsing → Chunking → Embedding → Vector Indexing

Phase 2: Runtime (Query)

User Query → Query Embedding → Semantic Search → Context Augmentation → LLM Response

API Operations

Operation	Purpose
Retrieve	Returns raw source chunks (custom pipelines)
RetrieveAndGenerate	Full RAG with citations
GenerateQuery	Natural language to SQL

---

Vector Store Options

Store	Key Characteristics
OpenSearch Serverless	Default, zero setup, serverless
Pinecone	High performance, purpose-built
Neptune	Graph-based, relationship-aware
Aurora RDS	SQL compatibility
Redis Enterprise	Low latency, real-time
MongoDB Atlas	Flexible schema

---

Chunking Strategies

Comparison

Strategy	Best For	Trade-off
Default (~300 tokens)	Quick setup	May not optimize for content
Fixed-Size	Uniform content	Configurable overlap
Hierarchical	Large documents	Fewer results, more context
Semantic	Complex relationships	Additional FM costs
No Chunking	Pre-processed files	No page-level citations

Hierarchical Chunking Details

• Parent chunks (large context) + Child chunks (precise embeddings)
• Child retrieved first, then replaced with parent
• Not recommended with S3 vector bucket

Semantic Chunking Details

• Divides by meaningful boundaries
• Buffer size of 1 = 3 sentences (previous, current, next)
• Larger buffers capture more context but may introduce noise

---

Data Source Connectors

Connector	Multimodal Support
Amazon S3	Yes
Custom Data Source	Yes
Confluence	No
SharePoint	No
Salesforce	No
Web Crawler	No

Critical: Multimodal content ONLY works with S3 and Custom connectors.

---

Embedding Models

Model	Dimensions	Vector Types
Titan Embeddings G1 - Text	1536	float32
Titan Text Embeddings V2	256/512/1024	float32, binary
Cohere Embed	1024	float32, binary

Binary vectors: Lower storage cost, reduced precision.

---

Advanced Parsing

Parser	Cost	Multimodal	Best For
Default	Free	No	Text extraction
Bedrock Data Automation	Per-page	Yes	PDF/images
Foundation Models	Per-token	Yes	Custom extraction

Warning: Once you select advanced parsing, it applies to ALL PDFs in data source.

---

Multimodal Capabilities

Processing Approaches

Approach	Converts To	Best For
Nova Multimodal Embeddings	Native vectors	Visual/audio similarity
Bedrock Data Automation	Text	Text-based search

Timestamp Handling: Your application must extract segments based on start/end metadata.

---

Reranking

Purpose: Reorder retrieved results by relevance.

Benefits:

• Fewer, more relevant results
• Decreased cost and latency
• More accurate responses

Limitation: Text only - no images, audio, or video.

---

3. Agents

Architecture Overview

Core Components

• Foundation Models: Decision engine
• Action Groups: Executable capabilities
• Knowledge Bases: Contextual augmentation

Runtime Process Flow

User Input → Pre-processing (optional) → Orchestration Loop → Post-processing (optional) → Response

Orchestration Loop:

1. FM generates rationale
2. FM predicts action group OR KB query
3. Execute action (Lambda or return control)
4. Generate observation
5. Re-augment prompt
6. Loop until complete

---

Action Groups

Two Definition Methods

Method	Best For
OpenAPI Schema	Explicit API mapping
Function Details	Simplified setup

Lambda Integration

• Max 11 API operations per action group
• ONE Lambda function per group
• Payload limited to 6 MB

Return Control Pattern

When you need control over execution:

1. Configure RETURN_CONTROL
2. Agent returns parameters with invocationId
3. Your app executes action
4. Send results back with same invocationId

Note: If returnControlInvocationResults included, inputText is ignored.

---

Multi-Agent Collaboration

Architecture

Supervisor Agent (orchestrator)
    ├── Collaborator Agent 1
    ├── Collaborator Agent 2
    └── Collaborator Agent N

Key: Hierarchical model with supervisor routing to domain specialists.

---

Memory System

Type	Persistence	Scope
Session Context	Session duration	Same sessionId
Memory Context	1-365 days	Same memoryId

Association Trigger: Memory associates when endSession=true OR idleSessionTimeout expires.

---

Code Interpretation

Capabilities: Generate, run, troubleshoot code in sandbox.

Limits:

• Max 5 files per request
• Total 10 MB file size
• 25 concurrent sessions per account

Regions: us-east-1, us-west-2, eu-central-1

---

Orchestration Strategies

Strategy	Use Case
Default (ReAct)	Standard workflows
Advanced Prompts	Modified templates
Custom Orchestration	Lambda-based complex logic

Four Base Prompt Templates

1. Pre-processing (disabled by default)
2. Orchestration
3. KB response generation
4. Post-processing (disabled by default)

Critical Warning: Agent instructions IGNORED if: single KB + default prompts + no action groups + user input disabled.

---

Session State

Type	Persistence
sessionAttributes	Full session
promptSessionAttributes	Single turn
conversationHistory	Multi-agent flows

---

4. Guardrails

How Guardrails Work

Processing Pipeline

User Input → Parallel Policy Evaluation → BLOCKED (no inference charges) OR PASSES → Model → Evaluation → Response

Key: If input blocked, model inference skipped = no charges.

---

Safeguard Types

Type	Purpose	Best For
Content Filters	Block harmful categories	General safety
Denied Topics	Block specific topics	Domain restrictions
Contextual Grounding	Detect hallucinations	RAG applications
Automated Reasoning	Validate logical rules	Regulated industries
Sensitive Info Filters	Block/mask PII	Privacy compliance
Word Filters	Block exact words	Profanity, terms

---

Content Filter Categories

• Hate
• Insults
• Sexual
• Violence
• Misconduct
• Prompt Attack

Filter Strength Configuration

• Start with Medium, adjust based on false positive rate
• Different thresholds for inputs vs outputs
• Actions: BLOCK or NONE (detect-only)

---

Denied Topics

Limits:

• Up to 30 topics per guardrail
• 200 chars (Classic) or 1,000 chars (Standard) per definition
• Up to 5 sample phrases per topic

Best Practices:

• Define topics crisply and precisely
• Don't include examples in definitions
• Don't define negative topics
• Use word filters for individual words

---

Contextual Grounding Check

Two Checks:

1. Grounding: Response accurate based on source?
2. Relevance: Response answers the query?

Limits:

• Source: 100,000 chars max
• Query: 1,000 chars max
• Response: 5,000 chars max
• Threshold: 0 to 0.99 (1.0 invalid)

Not suitable for: Conversational chatbots.

---

Automated Reasoning

Purpose: Validate against logical rules (HR policies, loan approvals, regulations).

Limitations:

• English (US) only
• No streaming API support
• Limited regional availability
• Cannot detect prompt injection

Variable Types: BOOL, INT, NUMBER, enum

---

Tier Comparison

Feature	Standard	Classic
Multi-language	Extensive	English, French, Spanish
Cross-Region	Supported	Not supported
Prompt Leakage Detection	Yes	No
Denied Topic Definition	1,000 chars	200 chars

Recommendation: Use Standard tier for new deployments.

---

Input Tagging (Required for Prompt Attack Detection)

<amazon-bedrock-guardrails-guardContent_xyz>
  User input goes here
</amazon-bedrock-guardrails-guardContent_xyz>

---

5. Model Customization

Customization Methods Overview

Method	Data Type	Purpose
Distillation	Prompts (unlabeled)	Transfer from teacher to student
Supervised Fine-Tuning	Labeled pairs	Task-specific accuracy
Continued Pre-Training	Unlabeled text	Domain knowledge expansion
Reinforcement Fine-Tuning	Prompts + reward function	Optimize for outcomes

---

Model Distillation

How It Works:

1. Provide input prompts (or use invocation logs)
2. Bedrock generates responses from teacher model
3. Student model fine-tuned on synthetic data

Two Paths:

• Path A: Upload prompts, Bedrock generates teacher responses
• Path B: Use existing CloudWatch Logs from production

Key: Can expand dataset up to 15,000 prompt-response pairs.

---

Supervised Fine-Tuning

Data Formats:

Non-conversational:

{"prompt": "Question?", "completion": "Answer."}

Conversational:

{
  "schemaVersion": "bedrock-conversation-2024",
  "system": [{"text": "You are helpful"}],
  "messages": [
    {"role": "user", "content": [{"text": "Q"}]},
    {"role": "assistant", "content": [{"text": "A"}]}
  ]
}

---

Continued Pre-Training

Data Format:

{"input": "Your domain text here..."}

Use Cases: Private documents, domain vocabulary, proprietary knowledge.

Supported Models: Titan Text G1 - Express only.

---

Reinforcement Fine-Tuning (RFT)

Three Stages:

1. Response Generation (4 responses per prompt)
2. Reward Computation (Lambda scoring)
3. Actor Training (GRPO optimization)

Two Approaches:

Approach	Best For
RLVR (Verifiable Rewards)	Code, math, objective tasks
RLAIF (AI Feedback)	Subjective quality

---

Training Data Requirements

Model	Min Records	Max Records	Max Tokens
Claude 3 Haiku	-	10,000	4,096 input + 2,048 output
Llama 3.1/3.2	5	10,000	16,000 total
Titan Text	1,000	500,000	4,096

Token Estimation: Use 6 characters per token.

File Size Limits:

• Training: 1 GB (fine-tuning), 10 GB (continued pre-training)
• Validation: 100 MB

---

Hyperparameters (Key Differences)

Model	Epoch Default	Batch Size	Special Features
Claude 3 Haiku	2	32 (4-256)	Early stopping
Llama 3.1	5	1 (fixed)	Wide learning rate range
Nova	2	-	Warmup steps

Key Details:

• Claude has automatic early stopping (overfitting prevention)
• Llama batch size is LOCKED at 1
• Each epoch multiplies token processing cost

---

Provisioned Throughput

MANDATORY for all custom models.

Commitment	Discount
No commitment	None
1 month	Moderate
6 months	Maximum

---

Regional Constraints

Model Family	Distillation Region
Claude & Llama	US West (Oregon)
Nova	US East (N. Virginia)

Note: Nova models cannot be copied to other regions.

---

6. Flows & Prompt Management

Bedrock Flows

Architecture

Directed node-based workflows: Input → Processing Nodes → Output

Key Elements:

• Nodes: Processing units
• Connections: Data (solid) or Conditional (dotted)
• Expressions: JSONPath-based data extraction ($.data.field)

---

Node Types

Category	Node	Purpose
Logic	Input	Entry point (exactly one)
	Output	Return response (multiple allowed)
	Condition	Route by logic (first match wins)
	Iterator	Process arrays (sequential)
	DoWhile	Loop (max 10 iterations)
AI	Prompt	Model inference
	Agent	Agent orchestration (supports multi-turn)
	Knowledge Base	RAG queries
Data	S3 Storage/Retrieval	Read/write S3
	Lambda	Custom code
	Inline Code	Python in-flow (max 5 nodes)

Key Details:

• Iterator is SEQUENTIAL, not parallel
• Inline Code NOT supported in async execution
• S3 Retrieval requires UTF-8 encoding

---

Multi-Turn Invocation (Preview)

Agent nodes can pause for user input:

1. First call: No executionId
2. Agent pauses: INPUT_REQUIRED event
3. Resume: Call with executionId + response
4. Complete: SUCCESS event

---

When to Use Flows vs Agents vs Direct API

Approach	Best For
Direct API	Simple calls, max control
Agents	Autonomous reasoning, tool use
Flows	Deterministic pipelines, visual dev

Hybrid: Embed Agent nodes within Flows.

---

Prompt Management

Core Features

• Variables: {{variable_name}} placeholders
• Variants: Alternative configurations
• Versioning: Immutable snapshots

Workflow

Create → Add Variables → Configure Model → Create Variants → Test → Save Version → Deploy

---

Prompt Optimization

Bedrock can automatically rewrite prompts for better results.

Best Practice: Optimize in English only.

Supported Models: Claude 3.x, Nova, Llama 3.x, Mistral Large

---

Intelligent Prompt Routing

Dynamically routes requests between models in same family:

1. Analyzes prompt
2. Predicts quality for each model
3. Routes to best quality/cost combo

Limitations:

• English only
• Same model family only
• No application-specific learning

---

Prompt Engineering Best Practices

Structure

[Context/Background]
[Input Content]
[Task/Instruction]
[Output Format]

Techniques

Technique	When to Use
Zero-Shot	Simple tasks
Few-Shot	Complex output, calibration needed
Chain-of-Thought	Reasoning tasks

Model-Specific Formatting

Model	Format
Claude	<example></example> tags
Titan	User: {}\nBot:
Llama	Meta's prompting guide

---

Quick Reference Cards

API Decision Tree

Need inference?
├── Single call, simple → InvokeModel / Converse
├── Batch processing → CreateModelInvocationJob
├── Streaming → InvokeModelWithResponseStream / ConverseStream
└── Tool use → Converse API

Cost Optimization Strategies

1. Global cross-region: ~10% savings
2. Flex tier: Non-time-sensitive workloads
3. Prompt caching: Repeated long contexts
4. Batch inference: Large-scale processing
5. Binary vectors: Lower storage costs

Feature Compatibility Matrix

Feature	InvokeModel	Converse	Notes
Text Generation	Yes	Yes
Embeddings	Yes	No
Image Generation	Yes	No
Tool Use	No	Yes
Multi-turn	Manual	Built-in
Guardrails	Yes	Yes

Common Exam Scenarios

Scenario	Solution
Need RAG with citations	Knowledge Bases + RetrieveAndGenerate
Block harmful content	Guardrails with Content Filters
Detect hallucinations	Contextual Grounding Check
Autonomous task execution	Agents with Action Groups
Reduce model costs	Distillation (smaller student)
Domain-specific knowledge	Continued Pre-Training
Visual workflow	Bedrock Flows
Prompt versioning	Prompt Management

---

Generated from official AWS Bedrock documentation