Domain 3: Applications of Foundation Models - Revision Notes

Here are complete revision notes for the AWS Certified Generative AI Developer Professional (AIP-C01) exam, synthesized directly from the practice questions.

Topic 1: Amazon Bedrock Guardrails

This was the most frequent topic. Master every capability.

Core Capabilities:

• Content Filters: Block harmful categories (Hate, Insult, Sexual, Violence).
• Denied Topics: Custom definitions to block specific subjects (e.g., “Medical Advice”, “SQL Queries”).
• Word Filters: Block specific keywords/profanity.
• Sensitive Information Filters (PII): The only feature that redacts PII in real-time (in-flight) during the request/response cycle.
• Contextual Grounding: Detects hallucinations. Checks if the response is supported by the source data (RAG).
• Prompt Attack Filters: Specifically blocks Prompt Injection and Jailbreaks (e.g., “Ignore previous instructions”, SQL injection attempts).

Exam Cues & Scenarios:

• “Dynamic Filtering” / “Time-based rules”: Do NOT use APIs to update guardrails in real-time. Use Tags on guardrails + EventBridge to switch context.
• “Redact PII in real-time”: Guardrails. (Macie is for storage scanning, not real-time redaction).
• “Prevent SQL Injection”: Prompt Attack Filters (part of Guardrails).
• “Prevent Hallucinations” / “Factuality”: Contextual Grounding Check.

---

Topic 2: Security & Networking

Focus on private connectivity and organizational control.

Private Access (No Public Internet):

• The Solution: Interface VPC Endpoints (PrivateLink).
• The Symptom: If a Lambda in a private subnet times out calling Bedrock, it’s missing a VPC Endpoint (or NAT Gateway, but Endpoint is preferred for security).
• The Trap: Do not use Gateway Endpoints (only for S3/DynamoDB) or Client VPN (for humans, not services).

Organizational Governance:

• “Restrict Models across the Org”: Use Service Control Policies (SCPs).
  • Scenario: Block Marketplace models or allow only specific Model IDs for everyone.
  • Why: IAM is local to an account; SCPs rule the organization.
• “Encryption at Rest”:
  • To enforce encryption (e.g., must use KMS Customer Managed Keys): Use AWS Config Rules.
  • To audit storage after the fact: Use Amazon Macie.

---

Topic 3: Bedrock Agents & RAG

Understanding the difference between APIs is critical.

Knowledge Bases (RAG):

• RetrieveAndGenerate API: Does it all—Retrieves chunks, generates the answer, and provides citations. Use this for “least overhead.”
• Retrieve API: Only fetches the text chunks. You must handle generation yourself.

Agent Security:

• Defense Layers:
  1. System Prompts: Define the “persona” and scope.
  2. Pre-processing Prompt: An Agent step that validates input safety before execution.
  3. Guardrails: The hard enforcement layer.
• The Trap: AWS WAF protects against web attacks (DDoS, XSS), but it cannot detect prompt injection or semantic attacks.

---

Topic 4: Data Engineering for GenAI

Focus on “Least Operational Overhead.”

Pipelines:

• Unstructured to JSONL (Fine-tuning): Use AWS Glue (Crawler + ETL).
  • Why: It is serverless and handles schema/metadata (Data Catalog) automatically.
  • The Trap: Amazon EMR requires managing clusters (high overhead). Athena is for SQL queries, not heavy text transformation.
• Lineage & Auditing:
  • Prompt Lineage: Use Bedrock Prompt Management.
  • Data Lineage: Use Glue Data Catalog.
  • Audit Trail: S3 Server Access Logs + CloudTrail.

---

Topic 5: SageMaker AI Governance

Distinguish between the different “SageMaker” tools.

• SageMaker Clarify: Detects Bias (Pre-training and Post-training). Fairness evaluation.
• SageMaker Model Monitor: Detects Drift (Data drift, Model quality drift) over time.
• SageMaker Model Cards / Registry: Documentation. Stores metadata, intended use, risk ratings, and compliance info.
• FMEval: Open-source library for evaluating LLMs (often used with Clarify).