AIP-C01: Study Notes

← Back to Overview

Study Strategy

These notes cover all exam domains comprehensively. Use the Quick Refresher for last-minute review. Domain 1 has dedicated notes at Domain 1 →.

---

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

See dedicated notes: Domain 1: FM Integration & Data Management →

---

Domain 2: Implementation and Integration (~26%)

2.1 Agentic AI & Bedrock Agents

Overview: Amazon Bedrock Agents enable multi-step reasoning workflows by orchestrating FM calls, API actions, and knowledge base retrievals.

Key Concepts:

1. Action Groups

   • Definition: Lambda functions that Agents can invoke to interact with external systems
   • Use case: Querying a database, calling a REST API, writing to S3
   • Example: An HR Agent calls a Lambda to fetch employee leave balance

2. Knowledge Base Integration

   • Definition: RAG pipeline attached to an Agent to retrieve relevant context
   • Use case: Document Q&A, policy lookup, customer support
   • Example: Agent retrieves relevant docs from OpenSearch before answering

3. Orchestration Trace

   • Definition: Step-by-step trace of Agent reasoning and tool calls
   • Use case: Debugging agentic behavior
   • Example: enableTrace: true in InvokeAgent API call

---

2.2 RAG Architecture & Chunking Strategies

Overview: Retrieval-Augmented Generation combines FM inference with context retrieval from a vector store to ground responses in real data.

Chunking Strategies:

Strategy	Description	Best For
Fixed-size	Split by token count (e.g., 300 tokens)	Simple docs, consistent structure
Semantic	Split by topic/meaning	Long-form content, varied structure
Hierarchical	Parent + child chunk structure	Complex docs with sections

When to Use:

• Use Fixed-size when: documents have uniform structure
• Use Semantic when: documents vary widely in structure
• Use Hierarchical when: you need both broad context and fine-grained retrieval

---

2.3 API Integration Patterns

Decision Tree:

Which Bedrock API to call?
├─ Need a complete, synchronous response? → InvokeModel
├─ Need streaming / low-latency UX? → InvokeModelWithResponseStream
└─ Multi-step agentic workflow? → InvokeAgent

---

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

3.1 Guardrails for Amazon Bedrock

Overview: Guardrails apply content filters and safety controls to both inputs (prompts) and outputs (responses) from foundation models.

Step-by-Step: Creating a Guardrail

1. Define Content Filters

   • What: Configure sensitivity levels for harmful categories (hate, violence, sexual, insults)
   • Why: Prevent inappropriate content from being generated or passed through

2. Configure PII Redaction

   • What: Detect and mask/block PII (names, emails, SSNs, credit cards)
   • Why: Compliance with data privacy regulations

3. Set Denied Topics

   • What: Specify topics the FM should refuse to discuss (e.g., competitor products)
   • Why: Business policy enforcement

4. Apply to Inference

   • What: Pass guardrailIdentifier and guardrailVersion in API calls
   • Why: Guardrails are only active when explicitly applied

---

3.2 IAM & VPC Security

Service/Tool Comparison:

Method	Primary Use	Key Feature	When to Use
IAM Policies	Access control	Least-privilege resource access	Always — required baseline
VPC Endpoints (PrivateLink)	Private connectivity	No public internet routing	Compliance, data residency
Resource Policies	Cross-account access	Bedrock model access from other accounts	Multi-account architectures

---

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

4.1 Cost Optimization Strategies

Provisioned Throughput vs On-Demand:

✅ Use Provisioned Throughput when:

• Traffic is predictable and consistent
• You need guaranteed model units (MUs) available
• Running 24/7 workloads where PTU commitment is cheaper than on-demand

❌ Avoid Provisioned Throughput when:

• Traffic is sporadic or unpredictable
• Testing or development workloads
• Short-lived experiments

Token Efficiency Best Practices:

• ✅ DO: Keep system prompts concise — every token costs money
• ✅ DO: Use streaming to improve perceived latency without changing cost
• ✅ DO: Set maxTokens explicitly to prevent runaway responses
• ❌ DON'T: Repeat the full conversation history when only recent context is needed

---

4.2 Monitoring & Troubleshooting

Common Issues:

Issue	Cause	Solution
ThrottlingException	Exceeded on-demand TPS	Switch to Provisioned Throughput or reduce request rate
High latency	Synchronous InvokeModel	Switch to InvokeModelWithResponseStream
Poor RAG retrieval	Low relevance scores	Tune chunking strategy or embedding model
Guardrail blocking valid content	Filter sensitivity too high	Reduce filter strength or update denied topics

---

Quick Reference

Key Acronyms

Acronym	Full Form	Meaning
FM	Foundation Model	Large pre-trained AI model (Claude, Llama, Titan, etc.)
RAG	Retrieval-Augmented Generation	FM + context retrieval from a vector store
PTU	Provisioned Throughput Unit	Reserved model capacity for predictable performance
PII	Personally Identifiable Information	Data that can identify an individual
MU	Model Unit	Unit of Bedrock Provisioned Throughput

Important Bedrock Limits & Notes

Resource	Key Detail
Claude context window	Up to 200k tokens
PTU commitment period	1 month or 6 months
OpenSearch Serverless vector dimensions	Up to 16,000 dimensions
Bedrock Knowledge Base chunk size	Configurable (default ~300 tokens)

---

← Back to Overview | Quick Refresher → | Exam Tips →