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: Always prefer the AWS-managed, purpose-built service over a custom-built solution (Bedrock Knowledge Bases > custom vector store scripts).
- Cost-aware: Questions asking for "most cost-effective" favor 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 (VPC Endpoints), least-privilege IAM, and Guardrails over post-hoc security additions.
- RAG over fine-tuning: When knowledge is dynamic or traceability matters, RAG is almost always the right answer over fine-tuning.
Keyword Detection Table
| If you see... | Look for this in the answer... |
|---|---|
| "Large context window needed" | Claude (Anthropic) — up to 200k tokens |
| "Open-source / custom fine-tuning needed" | Llama (Meta) |
| "Efficient inference at lower cost" | Mistral |
| "AWS-native embeddings for RAG" | Titan Text Embeddings |
| "Multilingual embeddings" | Cohere Embed |
| "Document Q&A / knowledge retrieval" | Knowledge Bases (RAG) |
| "Multi-step reasoning / tool use / external API calls" | Bedrock Agents + Action Groups |
| "Full synchronous response" | InvokeModel |
| "Streaming / low-latency UX / chat interface" | InvokeModelWithResponseStream |
| "Vector store for RAG" | Amazon OpenSearch Serverless (or Aurora pgvector) |
| "Content moderation / block harmful content" | Guardrails for Amazon Bedrock |
| "PII detection and redaction" | Guardrails — PII filter |
| "Predictable throughput / guaranteed no throttling" | Provisioned Throughput (PTU) |
| "Sporadic traffic / dev/test environment" | On-demand pricing |
| "Bulk, non-real-time inference jobs" | Batch inference |
| "Audit trail / compliance logging" | AWS CloudTrail |
| "Operational metrics / alarms / dashboards" | Amazon CloudWatch |
| "Private connectivity to Bedrock" | VPC Endpoint (PrivateLink) — bedrock-runtime |
| "Knowledge changes frequently / need source citations" | RAG (not fine-tuning) |
| "Detect hallucinations in RAG output" | Groundedness metric (Model Evaluation) |
| "Log all prompts and responses for AI governance" | Model Invocation Logging (to S3/CloudWatch Logs) |
Exam Traps
Look out for these!
RAG vs. Fine-tuning: RAG is the right answer when knowledge changes frequently or you need source attribution. Fine-tuning is for style and format adaptation on stable data. The exam frequently presents fine-tuning as a tempting distractor — default to RAG unless the scenario explicitly requires style or format changes.
Guardrails scope: Guardrails filter BOTH inputs AND outputs. A common wrong answer claims "Guardrails only filter model responses." They evaluate the prompt AND the response.
Guardrails must be explicitly applied: Guardrails do NOT automatically apply to all Bedrock calls. You must pass
guardrailIdentifier+guardrailVersionin each API request where you want them active.OpenSearch Serverless vs. managed OpenSearch: Bedrock Knowledge Bases use OpenSearch Serverless — not a standard managed OpenSearch cluster. The exam presents both options. Choose Serverless for Knowledge Bases.
InvokeModel vs. InvokeModelWithResponseStream: Both call the FM and cost the same in tokens. The difference is delivery: synchronous (full response at once) vs. streaming (token by token). Streaming does NOT save money — it improves perceived latency.
CloudTrail vs. CloudWatch: CloudTrail = audit log of API calls (who called what and when). CloudWatch = operational metrics (latency, errors, token counts). The exam swaps these frequently.
PTU for sporadic traffic: Provisioned Throughput requires a 1- or 6-month commitment. It is NOT appropriate for development, testing, or unpredictable workloads. On-Demand is correct for variable traffic.
Bedrock does not train on customer data: Amazon Bedrock never uses customer prompts or data to train foundation models. This is a hard guarantee — mark any answer claiming otherwise as incorrect.
Knowledge Base sync: After updating S3 content, you must manually trigger a sync on the Knowledge Base. The vector store does NOT auto-update when S3 changes.
Decision Quick Reference
"Which FM?"
Large context window (long docs, multi-turn)? → Claude (Anthropic)
Open-source, need to fine-tune? → Llama (Meta)
Efficient, lower-cost inference? → Mistral
AWS-native, embeddings, summarization? → Titan (AWS)
Multilingual embeddings? → Cohere Embed"RAG or fine-tuning?"
Knowledge changes frequently? → RAG
Need source traceability / citations? → RAG
Need to change model style or format? → Fine-tuning
Data is stable, model adapts its behavior? → Fine-tuning"Knowledge Base or Bedrock Agents?"
Simple document Q&A? → Knowledge Base (RAG)
Multi-step reasoning + tool use + APIs? → Bedrock Agents
Both retrieval + actions in one workflow? → Bedrock Agents with attached Knowledge Base"Which API?"
Synchronous full response? → InvokeModel
Streaming / low-latency chat UX? → InvokeModelWithResponseStream
Multi-step agentic workflow? → InvokeAgent"Which vector store?"
Bedrock Knowledge Base, fully managed? → Amazon OpenSearch Serverless
Existing PostgreSQL infrastructure? → Aurora PostgreSQL with pgvector
Enterprise doc retrieval with NLP ranking? → Amazon Kendra"PTU or On-Demand?"
Predictable, 24/7 consistent traffic? → Provisioned Throughput (PTU)
Sporadic, unpredictable, or dev/test? → On-Demand
Non-real-time bulk processing? → Batch Inference"Monitoring or Compliance?"
Audit trail for who called Bedrock and when? → CloudTrail
Operational metrics (latency, errors)? → CloudWatch
Log prompt + response content? → Model Invocation Logging (S3 or CloudWatch Logs)Final Strategy
- Domain 1 is 31% of the exam — FM selection, RAG architecture, and compliance are the biggest investment. Know chunking strategies, vector store options, and the FM comparison table cold.
- Domain 2 is 26% — Know the three Bedrock APIs and the distinction between Agents and Knowledge Bases. Be able to pick the right architecture for any scenario.
- Domains 1 and 2 together are 57% — Mastering these two domains alone gets you most of the way there.
- For every scenario, ask: Is this about retrieval (RAG/Knowledge Base), reasoning + action (Agents), cost (PTU/on-demand), safety (Guardrails), or compliance (CloudTrail/IAM)?
- Eliminate answers that: use the wrong service for the scenario, suggest fine-tuning when RAG fits better, propose PTU for sporadic workloads, or confuse CloudTrail with CloudWatch.