AIP-C01: Exam Guide
How the Exam Wants You to Think
The AIP-C01 is a Professional-level exam for developers building GenAI applications on AWS. It values architectural reasoning, cost-aware design, and responsible AI governance.
Answer Philosophy
- Architecture first: Prefer the AWS-managed, purpose-built service over a custom-built solution.
- Cost-aware: On-demand for sporadic traffic, PTU for steady 24/7 workloads, and batch inference for non-urgent bulk jobs.
- Security by default: Prefer private connectivity, least-privilege IAM, and Guardrails.
- RAG over fine-tuning: When knowledge is dynamic or traceability matters, RAG is almost always the right answer.
Keyword Detection Table
| If you see... | Look for this in the answer... |
|---|---|
| "Large context window needed" | Claude (Anthropic) |
| "Open-source / custom fine-tuning needed" | Llama (Meta) |
| "Efficient inference at lower cost" | Mistral |
| "AWS-native embeddings for RAG" | Titan Text Embeddings |
| "Generate summaries / conversational text with an AWS-native model" | Amazon Titan Text |
| "Convert text to vectors for semantic search" | Amazon Titan Embeddings |
| "Analyze images/video with low ops and summarize trends in dashboards" | Bedrock multimodal FM + Step Functions + QuickSight |
| "Document Q&A / knowledge retrieval" | Knowledge Bases (RAG) |
| "Multi-step reasoning / tool use / external API calls" | Bedrock Agents + Action Groups |
| "Unified message-based API across models" | Converse API |
| "Full synchronous response" | InvokeModel |
| "Model-specific request JSON / schema varies by FM" | InvokeModel |
| "Streaming / low-latency UX / chat interface" | InvokeModelWithResponseStream |
| "Persistent bidirectional real-time connection / server push" | API Gateway WebSocket API |
| "Vector store for RAG" | Amazon OpenSearch Serverless |
| "Store embeddings + perform low-latency similarity / k-NN search" | Amazon OpenSearch Service / Serverless Vector Engine |
| "Validate structured data quality before RAG / FM ingestion" | AWS Glue Data Quality |
| "Fully managed RAG with minimal custom pipeline code" | Bedrock Knowledge Bases |
| "Peak-period throttling, same FM, keep Bedrock API compatibility, low ops overhead" | Cross-Region inference profile |
| "Content moderation / block harmful content" | Guardrails for Amazon Bedrock |
| "Block a business-prohibited subject like investment advice" | Guardrails Denied Topics |
| "SageMaker LLM endpoint underutilized: short real sequence lengths, low concurrency, too many GPUs" | Reduce max sequence length + adjust tensor parallelism |
| "Predictable throughput / guaranteed no throttling" | Provisioned Throughput (PTU) |
| "Bulk, non-real-time inference jobs" | Batch inference |
| "Audit trail / compliance logging" | AWS CloudTrail |
| "HIPAA + full control over encryption keys" | Customer-managed KMS keys + CloudTrail |
| "Operational metrics / alarms / dashboards" | Amazon CloudWatch |
| "Private connectivity to Bedrock" | VPC Endpoint (PrivateLink) — bedrock-runtime |
| "Knowledge changes frequently / need source citations" | RAG |
| "Detect hallucinations in RAG output" | Groundedness metric |
| "Sequential evaluation stages with approval gates before model promotion" | Step Functions + Bedrock Model Evaluation / A-B comparison |
| "Log all prompts and responses for AI governance" | Model Invocation Logging |
| "FM inference may exceed 15 minutes" | ECS / Bedrock Batch — not Lambda |
| "API response time could exceed 29 seconds" | Async submit-and-poll — not synchronous API Gateway |
| "Human approval needed for hours or days" | Step Functions task token (Standard Workflow) |
| "Audit which S3 buckets contain PII before RAG ingestion" | Amazon Macie |
| "150-page document needs to fit in one prompt" | Claude (200k token context window) |
Exam Traps
Look out for these!
- RAG vs. Fine-tuning: RAG is right when knowledge changes frequently or you need source attribution.
- Guardrails scope: Guardrails filter BOTH inputs AND outputs.
- Guardrails must be explicitly applied: They do NOT apply automatically to all Bedrock calls.
- Guardrails vs. Knowledge Bases: Knowledge Bases improve relevance with retrieval; Guardrails enforce safety policy.
- Denied topics vs. word filters: denied topics block a subject semantically; word filters block exact words and can over-block.
- OpenSearch Serverless vs. managed OpenSearch: Knowledge Bases use OpenSearch Serverless.
- Converse vs. InvokeAgent: Converse is a unified model interaction API, not full agent orchestration.
- InvokeModel vs. Converse: InvokeModel request bodies can vary by model family; Converse reduces that variation with a more consistent message-based format.
- InvokeModel vs. InvokeModelWithResponseStream: Streaming does NOT save money.
- CloudTrail vs. CloudWatch: CloudTrail = audit; CloudWatch = operations.
- PTU for sporadic traffic: PTU is NOT appropriate for development or unpredictable workloads.
- Bedrock does not train on customer data: mark any answer claiming otherwise as incorrect.
- Knowledge Base sync: After updating S3 content, you must manually trigger a sync.
Decision Quick Reference
"Which FM?"
text
Large context window? → Claude
Open-source fine-tune? → Llama
Efficient inference? → Mistral
Embeddings? → Titan"RAG or fine-tuning?"
text
Knowledge changes often? → RAG
Need source traceability? → RAG
Need new style/format? → Fine-tuning"Knowledge Base or Bedrock Agents?"
text
Simple document Q&A? → Knowledge Base
Multi-step + tool use? → Bedrock Agents"Which compute for Bedrock?"
text
Per-request, < 15 min? → Lambda
Long-running / > 15 min? → ECS / EKS
Bulk async, non-real-time? → Bedrock Batch Inference
Human approval, hours/days? → Step Functions (Standard)"Service limits to remember"
text
Lambda max execution → 15 minutes
API Gateway integration timeout → 29 seconds (hard limit)
Step Functions Standard max → 1 year
Step Functions Express max → 5 minutes
Claude context window → 200,000 tokens
Titan context window → shorter (varies)