AIP-C01 Study Guide: Four Diagnostic Questions That Sort Almost Every Generative AI Developer Scenario
The AWS Certified Generative AI Developer – Professional exam (AIP-C01) describes generative AI applications that are almost working. The answers are wrong in a bad way, the right document is never found, an assistant talks about things it should not, or the bill is too high. Each question has four plausible fixes, and three of them work on the wrong layer. This guide gives you four questions to ask of every scenario, in order, so you can find the layer first and the service second.
Study guide · AWS Certified Generative AI Developer – Professional (AIP-C01)

The exam
75 questions in 180 minutes: 65 scored and 10 unscored, all multiple choice or multiple response. The passing score is 750 on a scaled score of 100 to 1,000. Five domains: Foundation Model Integration, Data Management, and Compliance (31%), Implementation and Integration (26%), AI Safety, Security, and Governance (20%), Operational Efficiency and Optimization for GenAI Applications (12%), Testing, Validation, and Troubleshooting (11%).
Amazon Bedrock is the centre of gravity. Among the 405 AIP-C01 questions we wrote for this question bank, about half (49%) are tagged with Bedrock itself, and another 70 only with its parts (Knowledge Bases, Guardrails, Agents, AgentCore and Prompt Management), so two questions in three involve Bedrock. AWS Lambda comes next (11%), as the glue. About one question in five (19%) is a troubleshooting scenario, more than the 11% weight of the testing domain suggests, because diagnosis shows up in every domain.
Question 1: is the model missing knowledge, or behaving the wrong way?
This is the first split, and the exam tests it constantly. A model that does not know your return policy has a knowledge gap. A model that knows it but writes in the wrong tone or format has a behaviour problem. They need different fixes.
| Symptom | Usual fix |
|---|---|
| Answers miss facts that change often or are private to the company | RAG, for example a Bedrock knowledge base. Retraining would be out of date the next time the facts change. |
| Output format or instructions are not followed | Prompt engineering first: a clear system prompt, few-shot examples, structured output. Version prompts in Bedrock Prompt Management. |
| A lasting style, vocabulary or task skill that prompts cannot teach | Fine-tuning a custom model, then serving it through Provisioned Throughput or a custom model deployment. |
| The team must switch models quickly | The Converse API (one request format across models) plus runtime configuration, instead of provider-specific InvokeModel bodies. |
Question 2: did retrieval fail, or generation?
In a RAG application, a bad answer has two possible causes: the right passages were never retrieved, or they were retrieved and the model used them badly. Changing temperature, model size or prompt wording does nothing for the first. Our explanations for Domains 1 and 5 come back to this split again and again.
- Nothing comes back, or old content comes back: check ingestion. A knowledge base reflects source changes only after a sync (an ingestion job), so automate the sync; and deleted source files must lead to deleted vectors.
- Exact identifiers are missed (part numbers, error codes, names): use hybrid search, which combines keyword and vector matching (in Bedrock Knowledge Bases, only for vector stores that support it, such as OpenSearch Serverless or Aurora PostgreSQL).
- The right passage is found but ranked low: add a reranker or transform the query.
- Answers lose context at chunk edges: revisit chunk size and overlap, or use hierarchical or semantic chunking.
- Retrieval is fine but the answer invents numbers: that is a generation problem. Measure faithfulness to the retrieved source, and use contextual grounding checks in Guardrails.
Evaluation follows the same split. Bedrock's RAG evaluation scores retrieval and generation separately, and LLM-as-a-judge scales quality scoring, but its ratings should be calibrated against human judgement for subjective qualities such as brand voice.
Going deeper: why vector search misses part numbers
An embedding model turns a passage into a vector that captures its meaning, so "cancel my order" and "stop my purchase" land close together. That is exactly what you want for paraphrases and exactly what fails for identifiers. "ZX-4471" and "ZX-4417" can land almost on top of each other, because to an embedding model they mean nearly the same thing, and a rare code may not be represented precisely at all. Keyword ranking (BM25 and similar) works the other way: it scores documents by the exact terms they share with the query, and gives rare terms extra weight. Hybrid search runs both and merges the results, so one query gets paraphrase matching and exact matching together. It is also why changing the embedding model is not a small change: documents and queries must be embedded by the same model, so a switch means re-embedding and re-indexing the whole corpus.
Beginner tip: build the smallest knowledge base you can: three short text files in an S3 bucket, one of them containing a made-up product code. Sync it and ask a question about the code. Then add a fourth file with a new fact, do not sync, and ask about that fact. Then sync and ask again. Ten minutes of this makes "the model is fine, the index is stale" an answer you recognise at once. Delete the knowledge base and its vector store afterwards so nothing keeps running.
Question 3: where should this control live?
Domain 3 questions rarely ask whether to add a control; they ask where. Each layer catches something the others cannot, and the exam treats them as additive, not alternatives.
| Requirement | Layer |
|---|---|
| Never discuss topic X, however it is phrased | A Guardrails denied topic. Word filters only catch the listed words and phrases. |
| Keep personal data out of prompts and answers | Mask in the application before the call; Guardrails PII filters as a safety net on input and output. |
| Instructions hidden in a retrieved web page or uploaded file | Indirect prompt injection: treat retrieved text as untrusted data and scan it, in addition to the guardrail on the user's prompt. |
| An agent must not be able to do damage even if tricked | Least-privilege IAM roles for its tools, checks before execution, and human approval for irreversible actions. |
| Auditors want to see what a past response said | Bedrock model invocation logging, which captures prompts and responses only if it was enabled before the request. |
| Per-version documentation of a model | SageMaker model cards linked to the model registry version. |
Identity is a separate trap: end users of your application get credentials through Amazon Cognito; employees reaching AWS accounts use IAM Identity Center.
Question 4: are you paying for tokens, for capacity, or for waiting?
Domain 4 is only 12%, but its options are easy to confuse because each one changes a different thing.
- Too many tokens: shorten the prompt, cache repeated context (prompt caching) or repeated questions (semantic caching), and route simple requests to a smaller model (model cascading). Temperature and streaming do not change what you pay.
- Not enough capacity: cross-Region inference absorbs bursts with no commitment (choose a geographic profile when data must stay in a geography); Provisioned Throughput reserves capacity for steady load.
- Users feel it is slow: stream the response to cut time to first token; run independent model calls in parallel.
- Large jobs that can wait: batch inference, which trades latency for a lower price.
A study order by weight
- Domain 1 (days 1 to 10 of the 35-day study plan on this site): model selection, Converse, the customisation spectrum, vector stores, chunking and embeddings, knowledge base sync, prompt management.
- Domain 2: agents (the reason-act-observe loop, tools, MCP, memory), multi-agent routing, Step Functions orchestration, deployment options, API Gateway and EventBridge integration.
- Domain 3: Guardrails in depth, prompt injection, PII and Macie, private connectivity through VPC endpoints, model cards and audit logging.
- Domains 4 and 5 together: cost and latency levers, CloudWatch monitoring, evaluation jobs, regression tests for prompts, and troubleshooting drills.
Practise
Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.
- AIP-C01 exam page · All 465 AIP-C01 questions · Timed mock exam · 35-day study plan
- By domain: Foundation Model Integration, Data Management, and Compliance · Implementation and Integration · AI Safety, Security, and Governance · Operational Efficiency and Optimization for GenAI Applications · Testing, Validation, and Troubleshooting
Exam facts as published in the AWS exam guide for AIP-C01; service shares refer to our own practice bank. Bedrock features, model availability and Region support change often; check the current documentation and the official exam page.