Read this the night before or on exam morning. It assumes you’ve studied the notes. It’s here to refresh, not to teach.
Exam day
- 75 questions in 180 minutes: 65 scored and 10 unscored questions you can’t spot. Multiple choice and multiple response (two or more correct from five or more options; you must get all of them).
- Scaled score 100 to 1,000; you need 750. Compensatory scoring: you don’t have to pass each domain, only the exam overall.
- No penalty for guessing; unanswered counts as wrong. About 2.4 minutes per question. Flag and move on.
- Domain weights: foundation model integration, data management and compliance 31%; implementation and integration 26%; AI safety, security and governance 20%; operational efficiency and optimisation 12%; testing, validation and troubleshooting 11%.
- Out of scope: model development and training, advanced ML, data and feature engineering.
How to read an AIP-C01 question
Read the last sentence first, then find the constraint words.
| The question says | Pick |
|---|---|
| Least operational overhead, managed | Bedrock native feature (Knowledge Bases, Guardrails, Prompt Management, Agents) over custom code |
| Without redeploying or without code changes | Config in AppConfig or Parameter Store; Prompt Management versions; inference profile |
| Up-to-date or private data | RAG, not fine-tuning |
| Consistent style, format or domain language | Fine-tuning or a prompt with examples |
| Reduce hallucination | Grounding plus contextual grounding check, citations, lower temperature |
| Reduce cost | Smaller model, caching, token limits, batch inference, routing |
| Reduce latency | Streaming, smaller or latency-optimised model, prompt caching, parallel calls |
| Throttling at peak | Cross-Region inference, backoff, SQS buffer, quota increase, Provisioned Throughput if steady |
| Audit, who asked what | CloudTrail plus model invocation logging |
| No internet path | VPC interface endpoints (PrivateLink) |
Elimination: answers that require training a model when the question is about knowledge, answers that use IAM or Guardrails for the wrong job, and anything with manual steps usually lose. In choose two or three questions, the picks are different parts of one solution.
Bedrock facts worth recalling
| Topic | Fact |
|---|---|
| Runtime APIs | InvokeModel, InvokeModelWithResponseStream, Converse, ConverseStream, ApplyGuardrail, async invoke for long jobs |
| Converse | One message format across models; separate system; tool use via toolConfig, toolUse and toolResult; stateless, so resend history |
| VPC endpoints | bedrock, bedrock-runtime, bedrock-agent, bedrock-agent-runtime |
| Data handling | Prompts and outputs aren’t used to train base models; customisation produces a private model copy |
| Guardrails | Content filters (including prompt attack), denied topics, word filters, sensitive information (block or mask), contextual grounding, Automated Reasoning |
| Chunking | Default (~300 tokens), fixed-size, hierarchical, semantic, none, custom Lambda |
| Embeddings | Same model for index and query; Titan Text Embeddings V2: 256, 512 or 1024 dimensions |
| Retrieval | Retrieve (chunks) vs RetrieveAndGenerate (answer with citations); hybrid search; reranker; metadata filters |
| Vector stores | OpenSearch Serverless (quick create), OpenSearch, Aurora pgvector, S3 Vectors, Neptune Analytics, others |
| Prompt Management | Templates with variables, immutable versions, variants for comparison |
| Prompt Flows | Prompt, knowledge base, condition, iterator, collector, Lambda, agent nodes |
| Throughput | On-demand (quotas), Provisioned Throughput (model units; needed for custom models), batch (cheaper, async) |
| Prompt caching | Stable prefix first, variable content last; short TTL (about 5 minutes) |
| Cross-Region inference | Inference profile ID instead of model ID; geographic vs global; application profiles for cost tags |
| Logging | Invocation logging off by default; to S3 and/or CloudWatch Logs; contains sensitive text |
| Errors | 429 throttling: backoff. Validation: fix request. Timeout: retry. AccessDenied: IAM, SCP, endpoint policy |
| Evaluation | Automatic, LLM-as-a-judge, human, RAG (retrieval only or retrieve and generate) |
Where each service fits
| Need | Service |
|---|---|
| Entity, PII and toxicity detection in text | Comprehend |
| PII in S3 before ingestion | Macie |
| Documents and forms to text | Textract, Bedrock Data Automation |
| Speech to text | Transcribe |
| Data quality rules | Glue Data Quality |
| Model versioning and approval | SageMaker Model Registry |
| Model documentation | SageMaker Model Cards |
| Fine-grained data lake access | Lake Formation |
| Orchestration with retries, loops, human approval | Step Functions (task tokens) |
| Live config, feature flags, gradual rollout | AppConfig |
| Tracing across services | X-Ray |
| Query prompts and responses | CloudWatch Logs Insights |
| Spend per application | Application inference profiles plus cost allocation tags |
| MCP tools: simple vs heavy | Lambda vs ECS; AgentCore Gateway for existing APIs |
Defaults that are usually right
- RAG for knowledge, fine-tuning for behaviour, workflow when the path is known, agent when it isn’t.
- Layer defences: WAF and API Gateway, input filter, Guardrails, output validation, logging.
- Smallest model that meets the quality bar; route and cascade the rest.
- Cache stable prefixes and repeated answers; cap output tokens.
- Evaluate on your own dataset, gate releases on scores, canary new models.
- Stream for user experience, SQS for long or bursty work.