Architecture Patterns

1 · Basic RAG Pipeline Deep dive →

When to use Answer questions grounded in company documents, reduce hallucination, dynamic knowledge that changes frequently. The most tested pattern on the exam.

1 · Ingest

S3 → Bedrock KB

Docs uploaded to S3

2 · Chunk

Split docs

Fixed / hierarchical / semantic

3 · Embed

Titan Embeddings

Chunks → vectors

4 · Store

Vector DB

OpenSearch / Aurora / managed

5 · Query

Top-k retrieval

User question → vectors

6 · Generate

Bedrock FM

Grounded response + citations

AWS services

Bedrock Knowledge Bases · S3 · OpenSearch Service · Titan Embeddings · Bedrock FM

Exam angle

Know every step. Know what can go wrong at each step. Know the chunking strategy tradeoffs (fixed / hierarchical / semantic). Know that "dynamic knowledge + grounded responses + citations" = this pattern, not fine-tuning.

2 · Advanced RAG with Hybrid Search + Reranking Deep dive →

When to use Basic RAG retrieval quality isn't good enough. You need better precision, especially with exact-term queries (SKUs, function names, error codes).

Query

User question

Raw text

Parallel A

BM25 keyword

Exact term match

Parallel B

Vector search

Semantic similarity

Fuse

Score fusion

Combine & normalize

Rerank

Bedrock reranker

Relevance re-score

Generate

Top reranked + FM

Answer with citations

Key additions vs. Basic RAG

OpenSearch hybrid search · Bedrock reranker models · custom scoring logic for fusion

Exam angle

Know when hybrid beats pure semantic: short queries, exact terminology, domain-specific jargon, names/codes. Pure semantic wins for long conceptual questions.

3 · Agentic Workflow (ReAct) Deep dive →

When to use Tasks requiring multi-step reasoning, tool use, and autonomous decision-making.

User request

Bedrock Agent receives

2 · Reason

Agent analyzes

What's needed?

3 · Act

Call tool

Action Group (Lambda)

4 · Observe

Process result

Feedback in context

5 · Loop

Done or iterate

Step Functions stop conditions

Final response

Guardrails applied

AWS services

Bedrock Agents · Action Groups (Lambda-backed) · Knowledge Bases · Guardrails · Step Functions (stopping conditions / timeouts)

Exam angle

Bedrock Agents (managed) vs. Strands Agents (open-source, more customizable). Exam prefers managed → Bedrock Agents. Know how to add human-in-the-loop via Step Functions callback pattern.

4 · Multi-Agent Orchestration Deep dive →

When to use Complex tasks that benefit from specialized agents working together. A research agent, a writing agent, a reviewer.

User request

To supervisor agent

2 · Route

Supervisor classifies

Which specialist(s)?

3 · Dispatch

Research / Code / Writing

Each with own tools + KB

4 · Collect

Specialist outputs

Parallel or sequential

5 · Synthesize

Supervisor aggregates

Final response

AWS services

AWS Agent Squad · Strands Agents · Bedrock Agents · Step Functions for orchestration

Exam angle

Multi-agent is appropriate for complex multi-domain tasks; overkill for single-purpose tasks where one agent suffices. If a question hints at overlap/coordination between different expertise areas, multi-agent wins.

5 · Model Cascading (Cost Optimization) Deep dive →

When to use High-volume workloads where most queries are simple but some require powerful models. The primary cost-optimization pattern.

Request arrives

API Gateway

2 · Classify

Lambda classifier

Query complexity

3A · Simple

Small / cheap FM

Haiku / Titan Lite

3B · Complex

Large / powerful FM

Sonnet / Titan Premier

4 · QC

Quality check

Below threshold?

5 · Escalate

Retry with bigger

If cheap fails

AWS services

Lambda (classifier) · API Gateway · Bedrock (multiple models) · Step Functions (routing logic) · CloudWatch (quality metrics)

Exam angle

Any question that says "reduce costs while maintaining quality" points here. Differentiates from simple routing because of the escalation step (quality gate + retry).

6 · GenAI Gateway (Enterprise) Deep dive →

When to use Multiple teams/applications need FM access with centralized governance, observability, and cost tracking.

Application request

Any internal team

2 · AuthN/Z

API Gateway + Cognito

Identity + permissions

3 · Safety

Guardrails

Content + PII filter

4 · Route

Model selection

Per team / request type

5 · Throttle

Per-team limits

Prevent capacity hogging

6 · Audit

Log + attribute cost

CloudWatch + CloudTrail

AWS services

API Gateway · Lambda · Cognito · Bedrock Guardrails · CloudWatch · CloudTrail · Cost Explorer · AppConfig

Exam angle

Question mentions "multiple teams," "enterprise," "centralized control," "cost attribution," or "consistent governance" → GenAI Gateway.

7 · Event-Driven FM Processing Deep dive →

When to use React to business events with AI-generated responses. New support ticket → auto-draft response. Document uploaded → auto-summarize.

1 · Source

Event fires

S3 / SQS / DDB Stream / EventBridge

2 · Extract

Lambda triggered

Parse event data

3 · Invoke

Bedrock call

Data + prompt

4 · Store

DynamoDB / S3

Result persisted

5 · Notify

SNS / SQS / webhook

Forward to consumer

AWS services

EventBridge · Lambda · S3 Events · DynamoDB Streams · SQS · Bedrock · SNS

Exam angle

Decouple FM processing from user-facing requests. Async via SQS when FM latency is acceptable. If latency matters, use different pattern (streaming / sync).

8 · Streaming Chat Interface Deep dive →

When to use Interactive chat applications where users expect real-time token-by-token responses.

Client connects

WebSocket / HTTP stream

API Gateway WS

Persistent connection

Lambda / ECS

InvokeModelWithResponseStream

Bedrock streaming

Tokens emitted as generated

Tokens to client

Real-time display

AWS services

API Gateway (WebSocket API) · Lambda · Bedrock Streaming API (InvokeModelWithResponseStream) · DynamoDB (conversation history)

Exam angle

If the question describes a chat interface, streaming is the right answer. Know WebSocket (bidirectional, persistent) vs. SSE (one-way) vs. chunked HTTP tradeoffs.

9 · CI/CD for GenAI Applications Deep dive →

When to use Production GenAI deployments that need automated testing, safe rollouts, and rollback.

Code commit

CodePipeline triggered

2 · Build

CodeBuild

Lambdas + prompts + guardrail configs

3 · Test

Prompt regression

Golden dataset · guardrails · integration

4 · Scan

Security

Prompt injection · IAM validation

5 · Deploy

Canary → staging

Small %, watch metrics

6 · Eval

Bedrock Model Eval

Roll forward or rollback

AWS services

CodePipeline · CodeBuild · CloudFormation / CDK · Bedrock Model Evaluations · CloudWatch Alarms

Exam angle

GenAI CI/CD includes prompt regression testing and guardrail validation — steps that don't exist in traditional CI/CD. If asked what's unique to GenAI CI/CD, these are the answers.

10 · Defense-in-Depth for FM Safety Deep dive →

When to use Any production FM deployment. This should be the default security posture — layered, not all-or-nothing.

1 · Network VPC endpoints for Bedrock; no public internet traffic.

2 · Identity IAM scoped to specific models and actions; Cognito for end-user auth.

3 · Pre-processing Comprehend PII detection + Lambda sanitization before FM sees input.

4 · Model-level Bedrock Guardrails — topic denial, content filtering, PII redaction, grounding, prompt-attack.

5 · Post-processing Lambda validates output format, accuracy, safety.

6 · API API Gateway rate limiting; WAF for abuse prevention.

7 · Audit CloudTrail + Bedrock Invocation Logs for forensic traceability.

AWS services per layer

7 layers, 7 service groupings. Each one maps to a domain of concern — know which service owns which layer.

Exam angle

Your CISSP/CCSP instinct maps directly. But watch the over-engineering trap — the exam wants the right level for the scenario, not max security. Don't add Lambda@Edge if Guardrails handles it natively.

Pattern selection matrix

Keyword → pattern. Use this as a memory aid during the exam.

When you see…	Pattern
"answer questions from company docs"	1 · Basic RAG
"retrieval quality isn't good enough"	2 · Hybrid + Rerank
"agent that can take actions / call tools"	3 · ReAct Agent
"specialized agents working together"	4 · Multi-Agent
"reduce cost while maintaining quality"	5 · Model Cascading
"multiple teams, central governance"	6 · GenAI Gateway
"react to business events with AI"	7 · Event-Driven FM
"chat interface, real-time responses"	8 · Streaming Chat
"production deployment with testing"	9 · CI/CD for GenAI
"any production FM security question"	10 · Defense-in-Depth