AIP-C01 last-minute revision: exam traps, memory hooks, a...

In Part 1, I explained why the AWS Certified Generative AI Developer - Professional (AIP-C01) certification stands apart from other AWS certifications. This follow-up post is a concise, 30-60 minute pre-exam revision guide covering exam traps, memory hooks, and quick notes across all five domains.

Disclaimer: These notes are a quick revision companion only. They are not a substitute for thorough exam preparation. Always refer to official AWS documentation and the recommended courses listed at the end of this post for comprehensive preparation.

Domain 1: Foundation Model Integration, Data Management, and Compliance (31%)

Foundation Models (FMs): Large pre-trained transformer models available via Amazon Bedrock: AWS Nova, Claude (Anthropic), Llama (Meta), Amazon Titan (text, embeddings, image), Jurassic-2 (AI21 Labs), Stable Diffusion (Stability AI). Select FMs based on task, latency, cost, and token limits.

Fine-tuning vs RAG:

Fine-tuning adapts an FM to a specific use case with proprietary training data. Titan, Cohere, and Meta models support fine-tuning via Amazon Bedrock. Text models need labelled prompt-completion pairs; image models need Amazon Simple Storage Service (Amazon S3) paths linked to descriptions. Secure training data with Amazon Virtual Private Cloud (Amazon VPC) + AWS PrivateLink.
RAG provides dynamic, up-to-date knowledge through vector stores (Amazon OpenSearch Serverless, Amazon Aurora pgvector, Amazon MemoryDB, Amazon ElastiCache, MongoDB Atlas, Pinecone, Redis Enterprise Cloud).
🧠 Memory Hook: Fine-tune = "teach the model new tricks"; RAG = "give the model a cheat sheet"
⚠️ Exam Trap: Fine-tune for style/tone changes; RAG for dynamic, up-to-date knowledge

LoRA Adapters: Lightweight fine-tuning technique. Amazon SageMaker AI Model Registry stores adapter versions with rollback strategies.

Chunking Strategies: Fixed-size, Hierarchical (smaller child chunks for precision, larger parent chunks for context), Semantic (FM-based, breaks content by meaning not length). Chunk size affects retrieval precision vs context.

Hybrid Search: Combines keyword search + vector search. Amazon Bedrock reranker models re-score results for improved relevance.

Query Expansion and Decomposition: Amazon Bedrock query expansion broadens search; AWS Lambda query decomposition breaks complex queries into sub-queries; AWS Step Functions orchestrates multi-step retrieval.

Embedding Models: Amazon Titan Embeddings, Cohere Embed. Match embedding model to vector store dimensions.

Vector Store Optimization: Binary vectors (32x compression vs float32), FP16 (16-bit scalar quantization for HNSW). Amazon OpenSearch Service Hierarchical Indices route queries from small fast top-level index to detailed domain-specific indices.

Prompt Engineering: Prompt = Instructions + Context + Input data + Output indicator. Few-shot prompting (examples of desired outputs). Chain of Thought (CoT) forces step-by-step reasoning.

Prompt Caching: Reuse previously processed prompts to reduce cost and latency.

Amazon Bedrock Prompt Management: Create, evaluate, version, and share prompts across teams. Supports variables in reusable templates.

Data Governance: Data residency, encryption at rest (AWS Key Management Service (AWS KMS)), encryption in transit (Transport Layer Security (TLS) 1.2+).

Amazon Bedrock Data Automation (BDA): Extracts structured data from multimodal inputs (documents, images, videos, audio). Uses Blueprints to specify extraction fields. Output: JSON, CSV, markdown, HTML.

🧠 Memory Hook: BDA = "Swiss Army knife for document processing"

Amazon Transcribe: Speech-to-text with PII redaction, automatic language identification, custom vocabularies, and ML-powered toxicity detection.

Bedrock Cross-Region Inference: Provides resilient FM deployments across regions for fault tolerance.

Domain 2: Implementation and Integration (26%)

Bedrock Agents: Action Groups (Lambda functions) + Knowledge Bases + Prompt Templates + Session Management. Action Groups rely on OpenAPI (Swagger) schema uploaded to Amazon S3.

🧠 Memory Hook: Agent = "Brain (FM) + Hands (Action Groups) + Memory (Knowledge Bases)"

Model Context Protocol (MCP): Standardised interface (JSON-RPC 2.0 over HTTP or stdio) for agent-tool interactions. MCP servers via Lambda (stateless) or Amazon Elastic Container Service (Amazon ECS) (complex tools).

🧠 Memory Hook: MCP = "USB-C for AI agents, one plug fits all tools"

Agent Frameworks: Strands Agents, AWS Agent Squad, Amazon Bedrock AgentCore for autonomous systems with memory and state management.

Agent Memory: Short-term (chat history via Sessions and Events). Long-term (extracted insights, user preferences stored as Memory Records). AgentCore Memory provides scalable, serverless storage.

Multi-Agent Workflows: Orchestrator delegates subtasks to worker LLMs, Synthesizer combines results. Chain of Sequence (sequential) or Parallelisation (concurrent execution, voting).

🧠 Memory Hook: Multi-agent = "assembly line with a foreman (orchestrator) and workers"

Amazon Bedrock Flows: Multi-step workflow orchestration with visual builder or JSON. Chain models, prompts, and conditions.

Sync vs Async Inference: Sync for real-time (InvokeModel); async for batch/long-running (InvokeModelWithResponseStream). Amazon Simple Queue Service (Amazon SQS) for async patterns.

Step Functions: Complex multi-service workflows, human-in-the-loop, error handling, parallel processing.

⚠️ Exam Trap: Step Functions for complex orchestration; Bedrock Agents handle simple multi-step tasks automatically

API Patterns: REST (Amazon API Gateway), GraphQL (AWS AppSync with real-time subscriptions), WebSockets for streaming.

Resilience Patterns: Exponential Backoff for retries (AWS SDK built-in). Circuit Breaker pattern via Step Functions + Amazon DynamoDB. API Gateway rate limiting.

🧠 Memory Hook: Circuit Breaker = "fuse box that trips before the whole house burns down"

AWS Cloud Development Kit (AWS CDK) / AWS CloudFormation: IaC for deploying GenAI stacks across environments. One CDK app + Stage construct per environment. Explicit env (account + region) per stack. Separate AWS accounts per environment.

⚠️ Exam Trap: Omitting env triggers environment-agnostic synthesis, breaking context lookups
🧠 Memory Hook: "One blueprint, multiple construction sites"

Continuous Integration / Continuous Delivery or Deployment (CI/CD) + AWS CodeDeploy: Canary, blue/green, rolling deployments for Lambda and compute targets.

Configuration and Secrets Management:

AWS Systems Manager Parameter Store: Static config (endpoints, URLs, free at 4 KB)
AWS Secrets Manager: Credentials with automatic rotation
AWS AppConfig: Dynamic runtime config without redeployment (feature flags, guardrail thresholds)
⚠️ Exam Trap: "rotation" = Secrets Manager. "without redeploying" or "feature flags" = AWS AppConfig
🧠 Memory Hook: "Phone book, vault with auto-lock-change, remote control"

Human-in-the-Loop (HITL): AI drafts, human refines. Route uncertain cases based on confidence scores. Collect feedback via API Gateway, store in DynamoDB.

Amazon Q Family:

Amazon Q Developer: Code generation, security scans, IDE extensions
Amazon Q Business: Enterprise GenAI assistant with data connectors (Amazon S3, SharePoint, Slack, Salesforce)
Amazon Q Apps: No-code GenAI productivity apps using natural language

Amazon Q Developer Project Configuration:

Uses .amazonq/ directory at the project root
Key file: .amazonq/rules.md (or multiple .md files in .amazonq/rules/)
Rules provide project-specific context, coding standards, architecture patterns, and constraints to Amazon Q Developer
Rules are scoped to the project, not global. Keep them concise and actionable
🧠 Memory Hook: .amazonq/rules.md = "instruction manual you leave for your AI coding assistant"

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

Amazon Bedrock Guardrails: Content filters (hate, insults, sexual, violence), denied topics, word filters, PII detection/masking, contextual grounding check (prevents hallucinations by measuring response alignment with retrieved context).

🧠 Memory Hook: Guardrails = "bouncer at both doors" (input AND output filtering)

Defense-in-Depth for Content Safety: Amazon Comprehend pre-processing > Amazon Bedrock Guardrails > Lambda post-processing > API Gateway filtering. Includes threat detection for prompt injection, jailbreaks, and input sanitisation.

🧠 Memory Hook: Defense-in-depth = "multiple security checkpoints, not just one gate"

Hallucination Reduction: Amazon Bedrock Knowledge Bases for grounding, confidence scoring, JSON Schema for structured outputs.

Amazon VPC Endpoints + AWS PrivateLink: Keep Amazon Bedrock traffic private within your VPC. Essential for sensitive fine-tuning data.

AWS Identity and Access Management (IAM) + AWS IAM Identity Center: Centralised access management. IAM Access Analyzer validates policies for least privilege.

Service Control Policies (SCPs) + Resource Control Policies (RCPs): SCPs restrict what accounts can do; RCPs restrict resource access.

⚠️ Exam Trap: SCPs don't grant permissions, they only restrict

Additional Security Services:

Amazon Macie: Data security and DLP for Amazon S3
Amazon Cognito: User auth for web/mobile apps
AWS WAF: Web application firewall
AWS Encryption SDK: Client-side encryption

Responsible AI: Fairness, explainability, transparency, human oversight, privacy and security, safety, controllability, veracity and robustness, governance.

Amazon Comprehend: NLP for sentiment, entities, PII detection, custom classification and entity recognition.

🧠 Memory Hook: Comprehend = "reads and understands text like a human"

Governance and Compliance: SageMaker AI model cards for documentation. AWS Glue Data Catalog for data lineage. AWS CloudTrail audit logging. Continuous monitoring for misuse, drift, and bias.

Domain 4: Operational Efficiency and Optimization (12%)

Amazon CloudWatch GenAI Observability: Track latency, token usage (InputTokenCount, OutputTokenCount), errors, API invocation counts. Time to First Token (TTFT) for streaming latency. Amazon CloudWatch Synthetics for canary monitoring.

Bedrock CountTokens API: Free API to estimate prompt token count before invoking the model.

AWS X-Ray: End-to-end distributed tracing across API Gateway, Lambda, Amazon Bedrock, Knowledge Bases.

🧠 Memory Hook: X-Ray = "MRI for your application's request flow"

Provisioned Throughput vs On-Demand: Reserved capacity for consistent performance vs pay-per-use. Provisioning is associated with a specific model ARN.

Prompt Caching: Caches static prompt prefix (instructions, system prompt). Only dynamic content tokenised on subsequent calls.

Cost Optimisation: Right-size models, cache prompts, batch inference, monitor token usage. Context Pruning (limit RAG chunks, filter via metadata, summarise old chat history). AWS Cost Explorer and AWS Cost Anomaly Detection for tracking GenAI spend.

Dynamic Routing (Intelligent Prompt Routing): Built into Amazon Bedrock. Routes complex queries to larger models, simple queries to smaller/cheaper models.

🧠 Memory Hook: Dynamic Routing = "express lane for simple questions, full service for complex ones"

Non-deterministic Outputs: Temperature, top-p, top-k control randomness. Lower temperature = more deterministic.

🧠 Memory Hook: Temperature = "creativity dial". 0 = robot, 1 = poet

Amazon SageMaker Clarify: Detects bias by measuring imbalances across demographic groups. Bias metrics: Class Imbalance (CI), Difference in Proportions of Labels (DPL).

Amazon SageMaker Model Monitor: Alerts via CloudWatch on quality deviations and data drift.

Semantic Caching: Cache similar queries' results using result fingerprinting. Edge caching via Amazon CloudFront for reduced latency.

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

Model Evaluation: Amazon Bedrock Model Evaluation for accuracy, robustness, toxicity. A/B testing, canary testing, cost-performance analysis.

LLM-as-a-Judge: Use an LLM to evaluate another LLM's outputs. Bedrock Evaluation Jobs measure RAG performance against benchmarks or LLM judges.

RAG Evaluation Metrics: Correctness, Completeness, Helpfulness, Logical Coherence, Faithfulness (how well responses align with retrieved text).

ROUGE Metric: Measures overlap of units (words, n-grams) between generated text and ground truth for summarisation or translation tasks.

Agent Debugging: Trace agent reasoning steps, validate action group responses, check knowledge base retrieval.

Bedrock Agent Tracing: Trace types: PreProcessing, Orchestration, PostProcessing, Guardrail traces. Shows which knowledge bases were hit, how action groups were invoked, and errors encountered.

Amazon SageMaker Ground Truth: Data labelling service for creating high-quality training datasets.

Troubleshooting Patterns: Inconsistent outputs, agent failures, retrieval misses, latency spikes.

Context Window Overflow: Dynamic chunking, prompt design optimisation, truncation error analysis.

Retrieval System Troubleshooting: Embedding quality diagnostics, drift monitoring, vectorisation resolution.

Amazon Augmented AI (Amazon A2I): Human review/correction loops for quality assurance. Vital due to non-deterministic nature of GenAI.

Exam decision boundaries

rotation = AWS Secrets Manager, not Parameter Store
without redeploying or feature flags = AWS AppConfig
consistent deployments across environments = one AWS CDK app with Stages
grounding or hallucination prevention = Amazon Bedrock Guardrails contextual grounding check or RAG with Knowledge Bases
standardised agent-tool interface = MCP
bias detection or explainability = Amazon SageMaker Clarify
data drift or model quality monitoring = Amazon SageMaker Model Monitor
human review loop = Amazon A2I
speech-to-text = Amazon Transcribe
text extraction from documents = Amazon Textract
conversational chatbot interface = Amazon Lex
contact centre AI = Amazon Connect + Amazon Lex
real-time subscriptions or GraphQL = AWS AppSync
event-driven = Amazon EventBridge
private Amazon Bedrock traffic = Amazon VPC Endpoints + AWS PrivateLink
sensitive data discovery in Amazon S3 = Amazon Macie

Key AWS services quick reference

Amazon Bedrock Ecosystem: Amazon Bedrock, Bedrock Agents, Amazon Bedrock AgentCore, Bedrock Knowledge Bases, Amazon Bedrock Guardrails, Amazon Bedrock Flows, Amazon Bedrock Prompt Management, Amazon Bedrock Data Automation (BDA), Bedrock Cross-Region Inference, Bedrock Model Evaluation

Agentic AI: Strands Agents, AWS Agent Squad, Model Context Protocol (MCP)

Data Processing and AI/ML: Amazon Textract, Amazon Transcribe, Amazon Comprehend, Amazon Rekognition, Amazon Lex, Amazon Titan, Amazon SageMaker AI, Amazon SageMaker Clarify, Amazon SageMaker Ground Truth, Amazon SageMaker JumpStart, Amazon SageMaker Model Monitor, SageMaker AI Model Registry, Amazon SageMaker Neo, Amazon A2I

Amazon Q Family: Amazon Q Developer, Amazon Q Business, Amazon Q Apps

Search and Vector: Amazon OpenSearch Service, Amazon Kendra, Amazon Neptune

Integration and Compute: AWS Lambda, Amazon Elastic Compute Cloud (Amazon EC2), AWS Step Functions, Amazon API Gateway, AWS AppSync, Amazon EventBridge, Amazon DynamoDB, Amazon SQS, Amazon Simple Notification Service (Amazon SNS), Amazon AppFlow

Infrastructure and Deployment: AWS CDK, AWS CloudFormation, AWS CodePipeline + AWS CodeBuild + AWS CodeDeploy, AWS AppConfig, AWS Systems Manager Parameter Store

Security, Identity, and Compliance: IAM + IAM Identity Center, AWS KMS, AWS Secrets Manager, Amazon Macie, Amazon Cognito, AWS WAF, Amazon VPC + AWS PrivateLink

Storage: Amazon S3, Amazon Elastic Block Store (Amazon EBS), Amazon Elastic File System (Amazon EFS)

Monitoring and Observability: Amazon CloudWatch, AWS X-Ray, AWS CloudTrail, AWS Cost Explorer, AWS Cost Anomaly Detection, Amazon Managed Grafana

Recommended preparation sources

I also highly recommend reading the relevant AWS service FAQ pages. They provide deeper understanding of service capabilities, limitations, and best practices that frequently appear in exam questions.

All the best on your AIP-C01 journey, and happy GenAI building! 🚀

推荐订阅源

DEV Community