AWS Certified Generative AI Developer - Professional (AIP-C01) practice questions

Domain coverage 5 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%)

Sample questions 11 of 110 shown

1. Q1 Foundation Model Integration, Data Management, and Compliance (31%)

   Which Amazon Bedrock feature provides a fully managed Retrieval-Augmented Generation (RAG) workflow that ingests documents from Amazon S3, performs chunking and embedding, and writes vectors to a configurable vector database?

   1. Bedrock Agents
   2. Bedrock Knowledge Bases
   3. Bedrock Studio
   4. Bedrock Model Evaluation

   Bedrock Knowledge Bases is the managed RAG service. It connects to Amazon S3, automatically chunks content using the selected embedding model (Titan Embeddings or Cohere Embed), embeds and writes vectors to OpenSearch Serverless, Aurora pgvector, MongoDB Atlas, Pinecone, or Redis Enterprise. At inference time, it retrieves and injects context into the prompt.
2. Q2 Foundation Model Integration, Data Management, and Compliance (31%)

   A team is building a low-latency RAG chatbot expected to handle 5,000 queries per second with vector search latency under 100 ms. They want a fully managed, auto-scaling serverless option. Which Bedrock Knowledge Bases vector store should they choose?

   1. Amazon OpenSearch Serverless (vector engine)
   2. Amazon Aurora PostgreSQL with pgvector
   3. Amazon DocumentDB
   4. Amazon Neptune

   OpenSearch Serverless with vector engine is fully managed, auto-scales based on OCU usage, and is Bedrock Knowledge Bases' default high-throughput option. Aurora pgvector is a strong option but requires capacity planning and is provisioned.
3. Q3 Foundation Model Integration, Data Management, and Compliance (31%)

   Which Amazon Bedrock model family is best suited for highly multilingual generation tasks with strong performance across European and Asian languages and is an AWS first-party model?

   1. Anthropic Claude 3.5 Sonnet
   2. Meta Llama 3.1
   3. Amazon Nova Pro
   4. AI21 Jurassic-2

   Amazon Nova Pro is an AWS first-party multimodal foundation model with strong multilingual capabilities (200+ languages) and is Bedrock-exclusive. Claude 3.5 Sonnet is excellent at reasoning but owned by Anthropic; Llama 3.1 is Meta's open-source model; Jurassic-2 has been deprecated.
4. Q4 Foundation Model Integration, Data Management, and Compliance (31%)

   A developer wants to switch between Anthropic Claude, Meta Llama, and Amazon Nova using the same application code without rewriting message handling logic. Which API should they use?

   1. InvokeModel
   2. Converse
   3. RetrieveAndGenerate
   4. InvokeAgent

   The Bedrock Converse API provides a unified, model-agnostic message interface (system, user, assistant turns) with consistent tool-use semantics, eliminating the model-specific request/response format required by InvokeModel. RetrieveAndGenerate is part of Knowledge Bases, and InvokeAgent is for calling Bedrock Agents.
5. Q5 Foundation Model Integration, Data Management, and Compliance (31%)

   When configuring a Bedrock Knowledge Base, which chunking strategy preserves complete logical units such as paragraphs by using a model to identify natural breakpoints rather than a fixed token count?

   1. Fixed-size chunking
   2. No chunking
   3. Semantic chunking
   4. Hierarchical chunking

   Semantic chunking uses an embedding model to detect topic transitions and creates variable-length chunks at natural boundaries, preserving meaning. Fixed-size chunking truncates at token thresholds. Hierarchical chunking creates parent/child chunks at multiple granularities.
6. Q6 Foundation Model Integration, Data Management, and Compliance (31%)

   Which embedding model is AWS's first-party option for Bedrock Knowledge Bases that supports output dimensions of 1024, 512, and 256 for trade-offs between cost and/or quality?

   1. Titan Text Embeddings V2
   2. Cohere Embed English v3
   3. OpenAI text-embedding-3-large
   4. Amazon Comprehend Topic Modeler

   Amazon Titan Text Embeddings V2 supports flexible 256/512/1024 dimensions, enabling teams to trade off between vector store cost and retrieval quality. Cohere Embed v3 is also available as a third-party model. OpenAI is not on Bedrock. Comprehend does not produce vector embeddings.
7. Q7 Foundation Model Integration, Data Management, and Compliance (31%)

   A regulated financial services customer must ensure their Bedrock prompts and responses stay entirely within their own AWS account and are not used for AWS service improvement. What guarantee does Bedrock provide by default?

   1. Customer data is automatically used to train AWS proprietary models unless they opt out
   2. Customer prompts and responses are NOT used to train any underlying foundation model and are not shared with model providers
   3. Customer data is shared with model providers for quality monitoring
   4. Customers must purchase Provisioned Throughput to keep data private

   Amazon Bedrock contractually guarantees customer data (prompts, completions, embeddings, fine-tuning data) is not used to train any underlying foundation model and is not shared with third-party model providers. This is the default property in all on-demand and provisioned throughput modes.
8. Q8 Foundation Model Integration, Data Management, and Compliance (31%)

   A team has 10,000 unlabeled internal engineering documents and wants to adapt a Bedrock foundation model to better understand its domain vocabulary but has no supervised input/output pairs. Which Bedrock customization technique should they choose?

   1. Fine-tuning
   2. Continued pre-training
   3. RAG with Knowledge Bases
   4. Prompt engineering only

   Continued pre-training (domain adaptation) uses unlabeled domain text and updates model weights to learn domain vocabulary and style. Fine-tuning requires labeled JSONL prompt/completion pairs. RAG augments at inference time but does not change weights. Selective Bedrock models support continued pre-training.
9. Q9 Foundation Model Integration, Data Management, and Compliance (31%)

   After fine-tuning a foundation model in Bedrock, what REQUIRED step is needed to invoke the resulting custom model for inference?

   1. No additional action; on-demand inference is automatic
   2. Purchase Provisioned Throughput (model units) for the custom model
   3. Deploy the model to a SageMaker endpoint
   4. Re-import model weights through Bedrock Custom Model Import

   Fine-tuned and continued pre-trained models on Bedrock are only served through Provisioned Throughput — model units must be purchased (hourly, 1-month, or 6-month commitment). Custom models do not support on-demand pricing. They run within Bedrock; SageMaker hosting is not used.
10. Q10 Foundation Model Integration, Data Management, and Compliance (31%)

    A team needs to build a Bedrock Knowledge Base based on documents containing confidential personal data subject to GDPR. They require embedding, chunking, and inference to occur only in eu-central-1. Which configuration is correct?

    1. Create the Knowledge Base in us-east-1 because Bedrock RAG is global
    2. Create the Knowledge Base, embedding model, and vector store in eu-central-1, and invoke the model through the eu-central-1 Bedrock endpoint
    3. Bedrock Knowledge Bases automatically locks data residency to the S3 source bucket region
    4. Override the region configuration with a VPC endpoint

    Bedrock is regional. For GDPR data residency, the source S3 bucket, Knowledge Base, embedding model calls, and vector store must all be created in the required EU region, and inference must call the Bedrock endpoint in that region. Bedrock is not global.
11. Q11 Foundation Model Integration, Data Management, and Compliance (31%)

    An e-commerce company uses an Amazon Bedrock Agent to handle product returns. The Agent has access to two tools: GetOrderDetails and ProcessRefund. During testing, the Agent fails to handle refunds for orders older than 30 days. The Agent does not call ProcessRefund and instead apologizes to the user. Which troubleshooting step provides the necessary visibility into the Agent's decision logic?

    1. Check CloudWatch Metrics for ProcessRefund Lambda Invocations and Errors
    2. Review CloudTrail Logs for InvokeModel events and check the system prompt
    3. Analyze the Agent Trace with enableTrace=true and check the orchestrationTrace for the rationale step
    4. Use Bedrock Model Evaluation with tool-use accuracy metric

    To understand the Agent's specific reasoning for a decision, enabling and checking the Agent Trace provides visibility into the thought process. The orchestrationTrace shows the Agent's rationale and observations from tool returns, offering insight into dynamic reasoning.

Answer all 110 questions with instant feedback →

Reader reviews 145 ratings

• Passed AIP-C01 with flying colors thanks to these practice exams. The questions are harder than the real thing, which is exactly what you want.

  ★★★★★ Lisa G.

• Was on the fence about buying the AIP-C01 practice test, but man am I glad I did. Nailed my certification today.

  ★★★★★ Tyler M.

• Solid Amazon prep material. The AIP-C01 questions are well-written and the answer rationales are thorough.

  ★★★★★★ Madison T.

• Got 856 out of 900 on the AIP-C01 exam. This question bank was my primary study resource. Highly recommend.

  ★★★★★ Alyssa D.

• Passed AIP-C01 with 94%. If you’re serious about this certification, get this question bank and thank me later.

  ★★★★★ Hunter P.

• I was skeptical about paying for exam prep, but the AIP-C01 bank saved me. Covered everything I needed.

  ★★★★★★ Nathan R.

Frequently asked questions

Free study resources

Community-verified analysis of 39 questions from real test-taker discussions — 11 deep analyses and 0 FAQs.

More Amazon exams