The Ultimate AWS Certified Generative AI Developer (AIP-C01) Cheat Sheet

Passing the AWS Certified Generative AI Developer - Professional (AIP-C01) requires shifting from "prompt engineer" to AI Architect. The exam focuses on building production-grade systems using Amazon Bedrock and SageMaker, with a heavy emphasis on security, cost, and RAG architectures.

If you are preparing for the exam, use this deep-dive cheat sheet to master the high-weightage domains.

1. Foundation Model (FM) Orchestration

The exam tests your ability to select the right model and API for the job.

The Bedrock API Selection Logic

• InvokeModel: Standard request-response. Best for batch or simple one-off tasks.
• InvokeModelWithResponseStream: Use for Chatbots. It improves User Experience (UX) by streaming tokens as they are generated.
• Converse API: The Unified API. Use this to write model-agnostic code. It handles the message-passing structure for Claude, Llama, and Mistral without rewriting logic.
• Provisioned Throughput: Reserved capacity for high-traffic apps or when using Custom/Fine-tuned models.

Inference Parameters (The "Knobs")

• Temperature: Range (0-1).
  • Low (0.1): Deterministic (Coding, Legal).
  • High (0.8): Creative (Marketing, Brainstorming).
• Top-P / Top-K: Use Top-P (Nucleus Sampling) for more dynamic, natural language. Use Top-K to strictly limit the model's vocabulary.
• Stop Sequences: Tell the model when to stop (e.g., \n, User:, or </json>).

2. RAG & Knowledge Bases (The 30% Domain)

Retrieval-Augmented Generation (RAG) is how you give models access to your private data.

Ingestion & Chunking Strategies

• Fixed-size: Fast but can cut off sentences mid-thought.
• Hierarchical: Links "Child" chunks (for retrieval) to "Parent" chunks (for context). Great for complex PDFs.
• Semantic: Uses embeddings to find natural "breaks" in topic. Most accurate, but most expensive to process.

Vector Store Selection

3. Agents & Action Groups

Agents use ReAct (Reason + Act) logic to perform multi-step tasks.

• Action Groups: Defined by OpenAPI schemas and Lambda functions. This is how an agent "calls" an external API (e.g., "Check stock in ERP").
• Return of Control: A critical feature where the Agent pauses and asks the calling application to handle an action (like a human approval for a $1000 refund).
• Prompt Management: Use Bedrock's managed prompt templates to version control your prompts separately from your Lambda code.

4. Security & Responsible AI

AWS treats security as a "Hard Gate." If the architecture isn't secure, it's the wrong answer.

Bedrock Guardrails

• The "One-Stop-Shop": Use Guardrails to block PII (SSNs, emails), filter hate speech, and prevent "competitor mentions."
• Contextual Grounding: Specifically detects Hallucinations. It checks if the model's answer is actually supported by the RAG source data.
• Data Privacy: Data used in Bedrock is never used to train the base foundation models. This is a common "True/False" exam trap.

5. Evaluation & Optimization

• Model Evaluation:
  • Automatic: Uses ROUGE or BLEU scores for objective tasks (Summarization).
  • Human: Use SageMaker Ground Truth for subjective tasks (Brand Voice).
• Prompt Caching: Essential for long-form RAG. It caches the "context" so you don't pay for the same 50-page PDF tokens every time a user asks a follow-up question.
• Model Routing: An architecture where a "Router" Lambda sends easy questions to Claude Haiku ($) and hard ones to Claude Sonnet ($$$).

Exam Strategy Summary

1. Hallucination problem? Answer: RAG or Contextual Grounding.
2. Behavior/Format problem? Answer: Fine-tuning.
3. Budget/Cost problem? Answer: Prompt Caching or Model Routing.
4. Governance? Answer: Bedrock Guardrails.

If you found this helpful, check out my other deep-dives into AWS and AI Implementation!