Reference

Tables, rules, architectures, and services — all in one place.

12 tables covering the key "when X vs Y" exam decisions

FM Integration

Vector Store Comparison

The exam's favorite question pattern. Match the exam trigger phrase to the correct vector store.

D1: FM Integration

Vector Store Comparison

The exam's favorite question pattern. Match the exam trigger phrase to the correct vector store.

Vector Store	Best For	Hybrid Search?	Managed?	Key Exam TriggerExam
Bedrock KB (managed store)	Fastest path, zero infra	Via Bedrock	Fully managed	'simplest' or 'least operational overhead'
OpenSearch Serverless	Scale + analytics + hybrid	Yes (BM25 + k-NN)	Serverless	'hybrid search' or 'analytics on retrieval data'
Aurora PostgreSQL (pgvector)	Existing RDS + SQL queries	Manual (SQL + vector)	Managed RDS	'existing relational database' or 'SQL queries alongside vector'
Neptune (graph)	Knowledge graphs + relationships	No	Managed	'entity relationships' or 'graph-based retrieval'
Kendra GenAI Index	Enterprise retrieval with connectors + ACL	Yes (native)	Fully managed	'enterprise search' or 'high-accuracy retrieval' or 'respect source permissions'
DocumentDB	MongoDB-compatible, HNSW/IVFFlat, up to 2000-dim	No	Managed	'existing MongoDB' or 'document-oriented database'
DynamoDB (co-index)	Metadata layer alongside vector stores	N/A (metadata only)	Fully managed	'metadata filtering' alongside semantic search

FM Integration

Chunking Decision Tree

Select the right chunking strategy based on content type.

D1: FM Integration

Chunking Decision Tree

Select the right chunking strategy based on content type.

Content Type	Best Chunking	Why
Short FAQ answers	Sentence-level or small fixed-size	Each answer is self-contained
Technical manuals	Hierarchical (section -> subsection)	Preserves document structure
Legal contracts	Semantic (by clause/paragraph)	Clauses must stay intact
Chat transcripts	Fixed-size with overlap	Even content, overlap preserves context
Code documentation	Hierarchical (by class -> method)	Code structure matters

FM Integration

Embedding Model Comparison

Choose the right embedding model based on language and modality needs.

D1: FM Integration

Embedding Model Comparison

Choose the right embedding model based on language and modality needs.

Model	Dimensions	Multilingual?	Best For
Amazon Titan Text Embeddings v2	256/512/1024	Limited	Cost-effective, AWS-native
Cohere Embed v3	1024	Yes (100+ languages)	Multilingual, high accuracy
Amazon Nova Multimodal Embeddings	Variable	Yes	Text + image + video + audio crossmodal search

Implementation

Bedrock Agents vs AgentCore

Critical distinction: managed agent service vs deployment platform for any agent framework.

D2: Implementation

Bedrock Agents vs AgentCore

Critical distinction: managed agent service vs deployment platform for any agent framework.

AspectExam	Bedrock Agents	Bedrock AgentCore
What	Managed agent service	Agent deployment platform
Framework	AWS-native only	Any (Strands, CrewAI, LangGraph, custom)
Model	Bedrock models only	Any FM (Bedrock, OpenAI, self-hosted)
Deployment	Fully managed	You deploy to AgentCore Runtime
Use when	Simple agent, fast setup, AWS-native	Custom framework, multi-model, complex orchestration

Implementation

AgentCore 9 Services

Know all 9 AgentCore services and the exam trigger phrases that point to each.

D2: Implementation

AgentCore 9 Services

Know all 9 AgentCore services and the exam trigger phrases that point to each.

Service	Function	Exam TriggerExam
Runtime	Serverless deployment, session isolation, microVMs	'deploy agent' + 'scale' + 'isolate sessions'
Gateway	Transform APIs/Lambda/MCP into agent-ready tools	'connect agent to existing APIs' + 'tool integration'
Policy	Cedar-based action boundaries, natural language authoring	'control what agents can do' + 'governance'
Identity	Agent auth for AWS + third-party services	'agent needs to access external APIs securely'
Memory	Session + long-term + episodic memory	'maintain context across sessions' + 'learn from past'
Observability	CloudWatch dashboards, OpenTelemetry, quality metrics	'monitor agent behavior' + 'debug agent decisions'
Evaluations	Correctness, helpfulness, safety scoring	'measure agent quality' + 'evaluate before production'
Code Interpreter	Secure sandbox for code execution	'agent needs to run code' + 'generate visualizations'
Browser	Web-based workflow execution	'agent needs to interact with web pages'

Implementation

MCP Server Implementation Patterns

The exam explicitly tests MCP server implementation. Know the stateless vs stateful distinction.

D2: Implementation

MCP Server Implementation Patterns

The exam explicitly tests MCP server implementation. Know the stateless vs stateful distinction.

MCP Server Type	Implementation	When to Use	Exam TriggerExam
Stateless (lightweight)	Lambda function	Read-only queries, simple tools, no persistent state	'lightweight tool access'
Stateful (complex)	ECS container	DB connection pooling, transactions, persistent state	'complex tools' + 'persistent connections'

Implementation

Q Business vs Bedrock KB + Custom App

Know when to use the managed enterprise search solution vs building custom.

D2: Implementation

Q Business vs Bedrock KB + Custom App

Know when to use the managed enterprise search solution vs building custom.

Feature	Q Business	Bedrock KB + Custom App
Setup	Plug-and-play, 40+ connectors (SharePoint, Confluence, Salesforce, Google Drive, Jira)	Custom build required
ACL	Built-in -- respects source permissions	Must implement manually
UI	Provided out of the box	Build with Amplify or custom
Customization	Limited	Full control
Identity	IAM Identity Center for SSO	Cognito or custom auth
Use when	Enterprise internal search, quick deployment, existing permissions matter	Customer-facing, custom UX, complex retrieval logic

Implementation

Model Cascading Pattern

Route queries to different models based on complexity. Frequently tested.

D2: Implementation

Model Cascading Pattern

Route queries to different models based on complexity. Frequently tested.

Query Complexity	Route To	Characteristics
Simple (70% of traffic)	Nova Micro	Cheapest, fastest
Medium (20% of traffic)	Nova Pro	Balanced cost/quality
Complex (10% of traffic)	Claude Sonnet	Highest quality

FM Integration

Service-to-Data-Type Mapping

Memorize which AWS service handles which data type in GenAI pipelines.

D1: FM Integration

Service-to-Data-Type Mapping

Memorize which AWS service handles which data type in GenAI pipelines.

Data Type	AWS Service	Use
Text	Comprehend	Entity extraction, sentiment, intent
Audio	Transcribe	Speech-to-text
Documents (PDF, images)	Textract	OCR, text extraction, table extraction
Images/Video	Rekognition / Bedrock multimodal	Object detection / FM analysis
Tabular data	Glue	ETL, data quality validation
Mixed/multimodal	Bedrock Data Automation	Automated processing pipeline

Safety & Security

Guardrails Filter Types

Know all 8 filter types, their configuration options, and what action they take.

D3: Safety & Security

Guardrails Filter Types

Know all 8 filter types, their configuration options, and what action they take.

Filter	Config	Action
Content filters	6 categories x 4 strengths (NONE/LOW/MED/HIGH)	Block harmful input/output
Prompt Attack	Separate from content filters	Detect jailbreaks, prompt injection
Denied topics	Custom topic + example phrases	Block specific subjects
Word filters	Custom word list + profanity toggle	Exact-match blocking
PII (sensitive info)	Per-entity type: BLOCK or ANONYMIZE	Mask or reject PII
Custom regex	Pattern-based detection	Org-specific identifiers
Contextual grounding	Grounding threshold + relevance threshold	Detect hallucinations
Automated Reasoning	Formal logic verification rules	Mathematically verify facts

Optimization

Caching Comparison

Critical exam distinction -- know all three caching layers and when to use each.

D4: Optimization

Caching Comparison

Critical exam distinction -- know all three caching layers and when to use each.

Method	How It Works	When to Use	Cost Impact
Bedrock Prompt Caching	Caches system prompt at Bedrock API level. 5-min TTL (1.25x write, 0.1x read) or 1-hour TTL (2.0x write, 0.1x read). Min 1024 tokens.	Same long system prompt across many requests	Up to 90% reduction on cached tokens
Semantic Caching (ElastiCache/DynamoDB)	App-level: embed incoming query, compare against cached query embeddings via cosine similarity	Repeated similar user queries	Eliminates FM call entirely (86% cost reduction with ElastiCache)
Exact-match Caching (ElastiCache)	App-level: hash identical queries, return cached response	Identical queries	Eliminates FM call entirely
Intelligent Prompt Routing	Auto-analyzes each prompt and routes to most appropriate FM based on complexity	Mixed-complexity workloads	Automated model cascading without manual routing logic

Testing

Evaluation Approaches

Match what you are evaluating to the correct method and AWS service.

D5: Testing

Evaluation Approaches

Match what you are evaluating to the correct method and AWS service.

What You're Evaluating	Method	Service
FM response quality	LLM-as-a-judge + human eval	Bedrock Evaluations
RAG retrieval relevance	Contextual grounding score	Bedrock Guardrails
RAG answer faithfulness	Grounding check against source	Bedrock Guardrails
Agent correctness	Automated eval on test cases	AgentCore Evaluations
Agent safety	Harmfulness scoring	AgentCore Evaluations
Production latency	End-to-end trace	X-Ray + CloudWatch
Prompt regression	Compare before/after on test set	Lambda + CloudWatch

35 trigger-to-answer mappings. When the exam says X, the answer is Y.

35 rules found

no code changes for model switchingAppConfig feature flags
NOT: Hardcoded model IDs
D1: FM Integration
hallucination detectionContextual grounding check
NOT: Content filter
D5: Testing
cross-region availabilityBedrock Cross-Region Inference
NOT: Manual multi-region deployment
D1: FM Integration
audit agent decisionsAgentCore Observability + CloudTrail
NOT: CloudWatch Logs alone
D2: Implementation
PII in code commentsGuardrails Standard tier
NOT: Basic PII filter
D3: Safety & Security
agent action boundariesAgentCore Policy (Cedar)
NOT: IAM policies alone
D2: Implementation
agent learns from past interactionsAgentCore Memory (episodic)
NOT: DynamoDB conversation history
D2: Implementation
connect existing APIs to agentAgentCore Gateway
NOT: Direct Lambda integration
D2: Implementation
deploy any framework agentAgentCore Runtime
NOT: Bedrock Agents
D2: Implementation
managed RAG, least overheadBedrock Knowledge Bases
NOT: Custom OpenSearch + Lambda pipeline
D1: FM Integration
custom fine-tuned model deploymentSageMaker + Model Registry
NOT: Bedrock Custom Model Training
D1: FM Integration
verify mathematical correctnessAutomated Reasoning checks
NOT: Contextual grounding
D3: Safety & Security
enterprise internal search with existing permissionsQ Business
NOT: Bedrock KB + custom app
D2: Implementation
developer code assistance in IDEQ Developer
NOT: Q Business
D2: Implementation
web interaction / form fillingAgentCore Browser
NOT: Lambda with headless browser
D2: Implementation
run code in sandboxAgentCore Code Interpreter
NOT: Lambda function
D2: Implementation
stateless lightweight MCP serverLambda
NOT: ECS
D2: Implementation
complex MCP server with persistent connectionsECS
NOT: Lambda
D2: Implementation
on-premises GenAIAWS Outposts
NOT: VPN to cloud Bedrock
D2: Implementation
edge low-latency GenAIAWS Wavelength
NOT: CloudFront
D2: Implementation
reduce cost of repeated system promptsBedrock Prompt Caching
NOT: Semantic caching
D4: Optimization
enterprise search, 40+ connectors, ACLQ Business
NOT: Kendra standalone
D2: Implementation
formal logic verificationAutomated Reasoning
NOT: Contextual grounding
D3: Safety & Security
GenAI chatbot with structured dialogueLex + Bedrock
NOT: Bedrock Agents alone
D2: Implementation
contact center AIConnect + Lex + Bedrock
NOT: Custom WebSocket + Bedrock
D2: Implementation
high-accuracy enterprise retrievalKendra GenAI Index
NOT: OpenSearch Serverless
D1: FM Integration
granular data access for GenAI dataLake Formation
NOT: IAM policies on S3
D3: Safety & Security
PII discovery in S3 before ingestionAmazon Macie
NOT: Guardrails PII filter
D3: Safety & Security
auto-select cheapest capable modelBedrock Intelligent Prompt Routing
NOT: Manual Lambda routing
D4: Optimization
non-real-time bulk processing at discountBatch inference (StartAsyncInvoke)
NOT: Provisioned throughput
D4: Optimization
hybrid search neededOpenSearch Serverless
NOT: Aurora pgvector
D1: FM Integration
existing relational database for vectorsAurora PostgreSQL (pgvector)
NOT: OpenSearch Serverless
D1: FM Integration
entity relationships / graph-based retrievalNeptune
NOT: OpenSearch
D1: FM Integration
existing MongoDB for vectorsDocumentDB
NOT: DynamoDB
D1: FM Integration
multilingual embeddings (100+ languages)Cohere Embed v3
NOT: Amazon Titan Embeddings v2
D1: FM Integration

4 reference architectures that combine all exam domains

Production RAG Pipeline

End-to-end RAG pipeline with security, caching, and monitoring. Combines Domain 1 (RAG), Domain 3 (Guardrails), and Domain 4 (caching).

API Gateway (Rate limit and authenticate incoming user request)

Lambda (Expand query using Bedrock (query reformulation))

Knowledge Base (OpenSearch) (Hybrid search (BM25 + k-NN) to retrieve relevant chunks)

Bedrock (Generate response with retrieved context as augmented prompt)

Guardrails (PII mask, grounding check, content filter on output)

API Gateway (Return filtered response to user)

Monitoring: CloudWatch (latency, tokens, cost) + X-Ray (trace) + CloudTrail (audit)

Cost: Semantic cache in DynamoDB for repeated queries

Multi-Agent System

Agent deployed on AgentCore with multiple tools, memory, and governance. Combines Domain 2 (agents, deployment) with Domain 3 (policy).

API Gateway (Receive user request and authenticate)

Strands Agent (AgentCore Runtime) (Plan and reason about how to fulfill the request)

AgentCore Gateway - Tool 1 (RAG lookup via Knowledge Base)

AgentCore Gateway - Tool 2 (CRM query via Lambda)

AgentCore Gateway - Tool 3 (Send email via SES + Lambda)

AgentCore Memory (Store session + episodic memory for learning)

AgentCore Policy (Enforce Cedar rules (e.g., 'cannot delete records'))

AgentCore Observability (Monitor in CloudWatch + OpenTelemetry)

Document Processing Pipeline

Multi-step document processing using Step Functions orchestration. Classic Domain 1 (data pipelines) pattern.

S3 (Document uploaded to S3 bucket)

EventBridge (Trigger Step Functions workflow on S3 event)

Textract (Step 1) (OCR -- extract text from scanned documents)

Bedrock Data Automation (Step 2) (Structured extraction from multimodal content)

Comprehend (Step 3) (Entity extraction (names, dates, amounts))

Bedrock + Guardrails (Step 4) (Summarize with PII anonymization)

DynamoDB + S3 (Step 5) (Store metadata in DynamoDB, processed output in S3)

Monitoring: Step Functions retry + catch + DLQ for error handling

Human-in-the-Loop Agent Workflow

Agent with confidence-based routing to human review. Combines Domain 2 (agents) with Domain 3 (governance).

API Gateway (Receive user request)

Strands Agent (Step 1) (Process request via AgentCore Runtime)

Agent Output (Step 2) (Generate action plan)

Lambda (Step 3) (Confidence check on action plan)

Auto-approve branch (If confidence > 0.9, execute automatically)

Wait-for-Callback branch (If confidence < 0.9, SNS notifies human reviewer)

Amplify Web UI (Human reviews and approves/rejects via callback token)

Execute (Step 4) (Execute approved action)

CloudTrail + Observability (Step 5) (Log decision for audit trail)

Monitoring: DLQ + SNS alert for errors

Reference

Vector Store Comparison

Vector Store Comparison

Chunking Decision Tree

Chunking Decision Tree

Embedding Model Comparison

Embedding Model Comparison

Bedrock Agents vs AgentCore

Bedrock Agents vs AgentCore

AgentCore 9 Services

AgentCore 9 Services

MCP Server Implementation Patterns

MCP Server Implementation Patterns

Q Business vs Bedrock KB + Custom App

Q Business vs Bedrock KB + Custom App

Model Cascading Pattern

Model Cascading Pattern

Service-to-Data-Type Mapping

Service-to-Data-Type Mapping

Guardrails Filter Types

Guardrails Filter Types

Caching Comparison

Caching Comparison

Evaluation Approaches

Evaluation Approaches

Production RAG Pipeline

Multi-Agent System

Document Processing Pipeline

Human-in-the-Loop Agent Workflow

Core AI/ML

AI Services

Compute

Storage/DB

Integration

Security

Monitoring

Data

Dev Tools

Customer Engagement

Visualization

Edge/Hybrid

Protocols