AWS Certified Generative AI Developer - Professional (AIP-C01) Practice Exam

What the AWS Certified Generative AI Developer - Professional (AIP-C01) exam covers

• Foundation Model Integration, Data Management, and Compliance233 questions
• Implementation and Integration195 questions
• AI Safety, Security, and Governance149 questions
• Operational Efficiency and Optimization for GenAI Applications91 questions
• Testing, Validation, and Troubleshooting83 questions

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

A sample of 10 questions with answers and explanations. Sign up free to practice all 751.

1. Question 1Foundation Model Integration, Data Management, and Compliance

   A team wants to test a revised wording of a production prompt template against real traffic before fully replacing the current wording. Which practice supports this safely?

   • ADisable the guardrail configuration while the new wording is evaluated
   • BOverwrite the existing production template file with the new wording
   • CSave the revision as a new version and compare it to the current oneCorrect
   • DIncrease the model's maximum token output limit before testing begins
   ✓ Correct answer: C

   Versioning preserves the known-good current wording while the revision is evaluated separately, so the application can roll back instantly if the new wording underperforms. Overwriting the live template removes that safety net, and token limits or guardrails are unrelated controls.

   Why the other options are wrong

   • ADisabling guardrails introduces unrelated content risk and does not help compare prompt wording quality.
   • BOverwriting the production template directly removes the ability to compare against or roll back to the prior wording.
   • DRaising the output token limit does not help evaluate whether a revised wording performs better.
2. Question 2Foundation Model Integration, Data Management, and Compliance

   Before associating a self-managed OpenSearch or Aurora PostgreSQL store with a Bedrock knowledge base, what must the team configure on the index or table?

   • AField mappings for the vector, text, and metadata fields the service expectsCorrect
   • BA fixed retention policy that automatically deletes any vectors older than 30 days
   • CA public endpoint that allows unauthenticated access from the Bedrock service
   • DA manual mapping of every possible user query to a pre-computed answer text
   ✓ Correct answer: A

   The knowledge base needs to know which index or table field holds the embedding vector, which holds the chunk's source text, and which holds metadata attributes, so these mappings must be defined and aligned with the embedding model's dimension before association succeeds.

   Why the other options are wrong

   • BThere is no requirement for an automatic 30-day retention policy on vectors; retention is managed through data source sync, not a time-based deletion rule.
   • CThe connection between Bedrock and the vector store uses authenticated access, not an open public endpoint.
   • DThe knowledge base retrieves relevant chunks dynamically; it does not rely on a precomputed query-to-answer mapping.
3. Question 3Foundation Model Integration, Data Management, and Compliance

   Before submitting a large training file for a model customization job, a team wants to catch formatting errors that would otherwise cause the job to fail partway through. What should they do first?

   • AConvert the training file into a single large JSON array before submitting
   • BSubmit the job immediately, since customization jobs auto-repair bad records
   • CReduce the training file size so fewer records are affected by an error
   • DValidate that every line is well-formed JSON matching the expected schemaCorrect
   ✓ Correct answer: D

   Running a schema validation pass over the JSON Lines file before submission identifies malformed or mismatched records early, avoiding a failed or partially completed customization job. Customization services do not automatically repair malformed records, converting to a single array changes the expected line-delimited structure rather than fixing errors, and simply shrinking the file does not address whether its records are correctly formatted.

   Why the other options are wrong

   • AThe expected format is JSON Lines, not a single nested array, so converting to an array does not fix formatting issues.
   • BCustomization jobs do not automatically repair malformed training records; a bad line can still cause a failure.
   • CReducing file size changes how many records are affected by an error but does not itself detect or prevent formatting errors.
4. Question 4Foundation Model Integration, Data Management, and Compliance

   A team is deciding how to store descriptive metadata for millions of chunks in a knowledge base. Storing large free-text metadata values directly alongside every vector increases index size and cost. What is a reasonable alternative?

   • ARemove all metadata entirely, including the attributes needed for filtering
   • BStore compact identifiers as metadata and look up full values separatelyCorrect
   • CStore the entire source document as a metadata value on every single chunk
   • DDuplicate every metadata value across every chunk regardless of relevance
   ✓ Correct answer: B

   Keeping only compact codes or identifiers in the vector index, with the full descriptive text held in a lighter-weight external store looked up when needed, keeps the index smaller and cheaper while still supporting attribute-based filtering. Storing entire documents or duplicating large values on every chunk multiplies storage cost, and removing metadata entirely eliminates the ability to filter at query time.

   Why the other options are wrong

   • ARemoving all metadata eliminates the ability to filter retrieval queries, which the team still needs.
   • CStoring an entire source document as metadata on every chunk multiplies storage far beyond what is needed for filtering.
   • DDuplicating large metadata values across every chunk regardless of relevance increases cost without added benefit.
5. Question 5Implementation and Integration

   A team wants to shrink the memory footprint of a very large OpenSearch vector index without changing the embedding model, by storing each vector component with fewer bits than full 32-bit floats. Which technique addresses this?

   • AResult reranking
   • BQuery term expansion
   • CVector quantizationCorrect
   • DMetadata field filtering
   ✓ Correct answer: C

   Techniques such as scalar or byte quantization store approximate, lower-precision versions of each vector component instead of full-precision floats, substantially shrinking index size and memory pressure. This typically introduces a small amount of additional approximation error on top of the ANN algorithm's own approximation.

   Why the other options are wrong

   • AReranking already-retrieved candidates does not affect how the base index stores vectors on disk.
   • BQuery expansion changes how the query is formed, not how vectors are stored.
   • DMetadata filtering narrows candidates by attribute values and does not reduce per-vector storage size.
6. Question 6Implementation and Integration

   A streaming response is interrupted partway through by an unrecoverable backend error. The client has already rendered several sentences of the answer. Which client-side behavior best matches good UX practice for this scenario?

   • AKeep the partial answer and offer retryCorrect
   • BRedirect the user to a generic error page
   • CRepeat the partial answer in a loop
   • DSilently discard the answer shown so far
   ✓ Correct answer: A

   Since useful content was already delivered, discarding it or hiding the failure wastes that value and confuses the user. Retaining the partial text and surfacing an explicit retry affordance lets the user decide whether to accept the partial answer or request it again.

   Why the other options are wrong

   • BRedirecting to a generic error page loses the partial content and gives no path to recover or retry the specific request.
   • CRepeating the partial answer in a loop does not resolve the failure and only confuses the user further.
   • DSilently discarding the answer shown so far throws away content the user already saw, offering nothing in its place.
7. Question 7AI Safety, Security, and Governance

   A company wants Lambda functions in a private subnet to call bedrock-runtime:InvokeModel without traversing the public internet. What should they configure?

   • AA NAT gateway with a route to the Bedrock public endpoint
   • BAn internet gateway attached to the private subnet
   • CA VPC gateway endpoint for the Bedrock Runtime service
   • DA VPC interface endpoint for the Bedrock Runtime serviceCorrect
   ✓ Correct answer: D

   Bedrock Runtime is accessed through AWS PrivateLink using a VPC interface endpoint, which places an elastic network interface with a private IP in the subnet so calls never leave the AWS network. Gateway endpoints only exist for S3 and DynamoDB, and NAT gateways or internet gateways both still route through the public internet.

   Why the other options are wrong

   • Aa NAT gateway still routes traffic over the public internet to reach Bedrock
   • Battaching an internet gateway exposes the subnet to the internet instead of keeping traffic private
   • Cgateway endpoints only support S3 and DynamoDB, not Bedrock
8. Question 8AI Safety, Security, and Governance

   While configuring a denied topic for competitor product comparisons, an administrator adds a few example phrases users might type. What is the purpose of these example phrases?

   • AThey are only used for documentation and have no effect on detection
   • BThey replace the natural-language topic definition entirely once they are added
   • CThey help the guardrail recognize variations of the topic during detectionCorrect
   • DThey are required so the guardrail can compute a contextual grounding score
   ✓ Correct answer: C

   Alongside the topic name and natural-language definition, an administrator can optionally supply example phrases that illustrate the topic. These examples give the underlying detection model more context, improving recognition of the topic across differently worded prompts.

   Why the other options are wrong

   • AExample phrases actively improve detection; they are not purely documentation.
   • BExample phrases supplement the definition; they do not replace it.
   • DGrounding scores come from the contextual grounding check, which is unrelated to denied topic examples.
9. Question 9Operational Efficiency and Optimization for GenAI Applications

   A conversational assistant built on Amazon Bedrock currently waits for the full model response before displaying any text, and users report that the assistant feels slow even though the total response time has not changed. Which change would most directly address this perception without altering the total generation time?

   • ASwitch to streaming so tokens display as they are generatedCorrect
   • BReduce the maximum number of tokens allowed in each response
   • CIncrease memory allocated to the Lambda function for parsing
   • DMove the vector store onto a higher performance search instance
   ✓ Correct answer: A

   Response streaming sends tokens to the client as soon as the model produces them, so the perceived wait before content appears drops dramatically even though the model still needs the same amount of time to finish the full response. This directly targets the felt experience of slowness without changing total generation time.

   Why the other options are wrong

   • BLowering the token limit shortens total generation time itself rather than changing how the existing response is perceived while it is produced.
   • CLambda memory affects compute for the function's own logic, not the pace at which the model streams or completes tokens.
   • DVector store performance affects retrieval latency for the knowledge base, not how response tokens are delivered to the client.
10. Question 10Testing, Validation, and Troubleshooting

    After confirming that fabricated answers stem from high sampling randomness combined with weak grounding instructions, which combination of changes best reduces hallucination for a fact-based RAG assistant?

    • ALower the temperature and require answers only from given contextCorrect
    • BSwitch to a larger context window without changing any instructions
    • CRaise the temperature and remove citation requirements from the prompt
    • DIncrease the output token limit and disable the retriever entirely
    ✓ Correct answer: A

    Lower temperature narrows sampling toward the most probable, context consistent tokens, while an explicit grounding instruction tells the model what to do when the context is insufficient, such as stating it does not know. Combined, these two changes directly target both causes identified in the diagnosis, randomness and weak grounding guidance.

    Why the other options are wrong

    • BA larger context window without instruction changes does not address either randomness or grounding discipline.
    • CRaising temperature and dropping citation requirements would push the model toward more, not less, ungrounded creativity.
    • DDisabling the retriever removes grounding context altogether, which would make hallucination more likely, not less.

Related AWS resources

• AWS Certified Generative AI Developer - Professional (AIP-C01) study guideKey concepts
• AWS practice examsAll AWS
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• AWS SCS-C03 practice examRelated
• AWS Certified AI Practitioner (AIF-C01) practice examRelated
• AWS CLF-C02 practice examRelated

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