Skip to content

AIP-C01: AWS Generative AI Developer Last-Minute Refresher

← Back to Overview

Final Review

This page is designed for the final "cram" session before stepping into the AIP-C01 exam.


Domain 1: FM Integration, Data Management, and Compliance (~31%)

FM Selection Quick Guide

FMVendorBest For
ClaudeAnthropicReasoning, large context (up to 200k tokens)
LlamaMetaOpen-source, fine-tuning
MistralMistral AIEfficient, high performance for size
TitanAWSEmbeddings, summarization, AWS-native

RAG Pipeline

Important Points:

  • S3 → chunking → embedding → OpenSearch Serverless (vector store)
  • Knowledge Bases manage this pipeline fully managed
  • pgvector on Aurora PostgreSQL = alternative vector store
  • Chunking options: fixed-size, semantic, hierarchical

Domain 2: Implementation and Integration (~26%)

API Quick Reference

Which API?
├─ Full synchronous response → InvokeModel
├─ Streaming / low latency → InvokeModelWithResponseStream
└─ Multi-step agentic workflow → InvokeAgent

Comparison Table

FeatureKnowledge Base (RAG)Bedrock Agents
Use CaseDocument Q&A, retrievalMulti-step workflows, tool use
FM + RetrievalYes (automatic)Yes (optional)
External API callsNoYes (via Action Groups + Lambda)
Best ForStatic knowledge lookupDynamic reasoning + actions

Domain 3: AI Safety, Security, and Governance (~20%)

Guardrails Quick Facts

  • Applies to both inputs and outputs
  • Features: content filters, PII redaction, denied topics, word filters
  • Must be explicitly applied via guardrailIdentifier in API calls
  • Does NOT automatically apply to all Bedrock calls

Security Decision Tree

Data must stay private / no public internet?
└─ Yes → VPC Endpoint (PrivateLink) for Bedrock

Need to restrict access to Bedrock models?
└─ Yes → IAM resource-based policies

Need to audit all FM API calls?
└─ Yes → AWS CloudTrail (not CloudWatch)

Need operational metrics (latency, errors)?
└─ Yes → Amazon CloudWatch

Domain 4: Operational Efficiency and Optimization (~12%)

Provisioned Throughput vs On-Demand

Provisioned ThroughputOn-Demand
TrafficPredictable, consistentSporadic, unpredictable
PricingFixed (per MU/hour)Pay per token
Commitment1 or 6 monthsNone
Best ForProduction 24/7Dev/test, bursts

Domain 5: Testing, Validation, and Troubleshooting (~11%)

Model Evaluation Metrics

MetricWhat It Measures
GroundednessIs the response grounded in the retrieved context?
RelevanceIs the response relevant to the user's question?
AccuracyIs the response factually correct?
FluencyIs the response well-written and natural?

Key Service Decision Rules

Question asks about...

  • "Predictable throughput" → Provisioned Throughput
  • "Content moderation / filter" → Guardrails for Amazon Bedrock
  • "PII detection/redaction" → Guardrails (PII filter)
  • "Vector storage for RAG" → OpenSearch Serverless (or Aurora pgvector)
  • "Multi-step reasoning / tool use" → Bedrock Agents
  • "Audit logging / compliance trail" → AWS CloudTrail
  • "Operational monitoring / dashboards" → Amazon CloudWatch
  • "Private FM access, no internet" → VPC Endpoint (PrivateLink)
  • "Low latency / streaming response" → InvokeModelWithResponseStream

Key Acronyms

AcronymFull FormQuick Definition
FMFoundation ModelPre-trained large AI model (Claude, Llama, etc.)
RAGRetrieval-Augmented GenerationFM + vector store retrieval
PTUProvisioned Throughput UnitReserved Bedrock model capacity
PIIPersonally Identifiable InformationData that identifies an individual
MUModel UnitUnit of Bedrock Provisioned Throughput

Common Exam Traps

Watch Out!

  • InvokeModel vs Streaming: InvokeModel = full synchronous; InvokeModelWithResponseStream = chunk-by-chunk
  • Guardrails scope: Guardrails filter BOTH inputs AND outputs — not just one
  • CloudTrail vs CloudWatch: CloudTrail = audit/compliance; CloudWatch = operational metrics
  • OpenSearch Serverless: The go-to for Knowledge Bases — not standard OpenSearch managed cluster
  • PTU timing: On-demand is better for sporadic traffic — don't default to PTU

Last 5 Minutes Before Exam

Must Remember

  1. Guardrails — apply to both inputs AND outputs; must be explicitly applied per call
  2. PTU — predictable traffic only; 1- or 6-month commitment
  3. RAG vector store — OpenSearch Serverless (primary), pgvector on Aurora (alternative)
  4. InvokeModelWithResponseStream — for low-latency streaming UX
  5. CloudTrail — required for compliance/audit; CloudWatch for ops monitoring

Quick Mental Check

  • Can you explain the difference between RAG and fine-tuning? ✓
  • Do you know when to use InvokeModel vs InvokeModelWithResponseStream? ✓
  • Can you describe what Guardrails filters cover? ✓
  • Do you understand when PTU is preferred over on-demand? ✓
  • Can you pick the right vector store for a Knowledge Base? ✓

You've Got This!

Take a deep breath. Review the decision trees. Trust your preparation. Good luck!

← Back to Overview | Study Notes | Exam Tips