15 Free AWS Certified Generative AI Developer – Professional (AIP-C01) Practice Questions with Answers (2026)

AppConfig is built for runtime configuration. It supports gradual deployment strategies and automatic rollback when a linked CloudWatch alarm fires, and functions pick up changes without being redeployed.

• B is wrong because updating environment variables changes every function's configuration at once, with no gradual rollout or automatic rollback.
• C is wrong because a new layer version requires every function to be updated and redeployed.
• D is wrong because overwriting an S3 object changes everything at once, with no staged rollout or alarm-based rollback.

Hybrid search combines keyword (text) matching with semantic vector search. Exact strings such as part numbers score highly on the keyword side, so the right sheet is retrieved, and OpenSearch Serverless supports hybrid search for knowledge bases.

• A is wrong because more dimensions do not make embeddings reliable at exact string matching.
• C is wrong because a reranker can only reorder what was retrieved; if the right sheet is not in the candidates, reranking cannot bring it back.
• D is wrong because temperature affects generation, not retrieval.

Metadata filtering restricts retrieval to documents whose metadata matches the filter. Because the backend sets the filter from the authenticated identity, not from anything the user types, another customer's chunks never reach the model.

• A is wrong because a prompt instruction is not an access control; the other customers' chunks would still be retrieved, and prompt injection could expose them.
• B is wrong because denied topics block subjects in conversation; they do not control which documents are retrieved.
• C is wrong because one knowledge base uses one embedding model, and embeddings do not provide access control anyway.

Prompt Management supports templates with variables and immutable versions. Applications that reference a specific version ARN only change when someone deliberately points them at a new version, so rollback is a reference change, and CloudTrail records the API activity.

• A is wrong because prompts are duplicated across repositories with no shared versioning or central audit.
• C is wrong because direct edits change prompts in place, so there are no immutable reviewed versions without building that yourself.
• D is wrong because a wiki has no versioned integration with applications and no audit of how prompts are used.

ConverseStream returns the response in chunks as tokens are generated. Pushing each chunk over the existing WebSocket connection lets the user start reading within moments.

• B is wrong because the user still waits for the full response.
• C is wrong because asynchronous email delivery is not an interactive chat.
• D is wrong because batch inference is for large offline jobs, not interactive chat.

Step Functions enforces the rule in code, outside the model. A task-token callback pauses the workflow until the reviewer responds, which can take days, and the execution history records every decision.

• A is wrong because instructions to a model are not enforcement; the model can be wrong or manipulated.
• B is wrong because denied topics block conversation subjects; they cannot pause a workflow or wait for a human.
• D is wrong because more reasoning steps still leaves the decision with the model, and adds no human approval.

A long-running container on ECS with Fargate can hold persistent connections and session state, which suits a complex, stateful MCP server. Lambda-based MCP servers fit lightweight, stateless tools.

• A is wrong because Lambda functions are stateless between invocations and stop after at most 15 minutes, so they cannot hold a 30-minute session or a persistent connection.
• B is wrong because a prompt description is not a tool; the agent cannot query live inventory through it.
• C is wrong because an export goes stale and a knowledge base cannot run live queries against the gateway.

Backoff with jitter spreads retries so they succeed instead of colliding, and a queue absorbs bursts for work that can wait. A US cross-Region inference profile routes requests across several US Regions, which adds throughput without committed capacity.

• C is wrong because more concurrent callers send more requests at once, which increases throttling.
• D is wrong because longer outputs use more tokens per request and make throttling worse.
• E is wrong because temperature controls randomness, not the cost or rate of a request.

The contextual grounding check scores each response for grounding (is it supported by the source passages?) and relevance (does it answer the query?), and blocks responses below the thresholds you set.

• A is wrong because you cannot list every possible unsupported claim in advance.
• B is wrong because a lower temperature reduces randomness but does not detect or block ungrounded answers.
• D is wrong because a denied topic blocks one subject; it does not check answers against the retrieved sources.

An interface VPC endpoint (AWS PrivateLink) keeps traffic on the AWS network. The endpoint policy limits what can be called through it, and the aws:SourceVpce condition denies Bedrock calls that do not arrive through that endpoint.

• B is wrong because a NAT gateway needs an internet gateway and sends traffic to the public Bedrock endpoint, which breaks the no-internet design.
• C is wrong because an API key changes authentication, not the network path, and long-term keys are not recommended for production.
• D is wrong because a gateway endpoint for S3 only carries S3 traffic.

Model invocation logging captures the request and response data for model invocations. Delivering it to S3 with Object Lock in compliance mode makes the records unchangeable for the retention period.

• A is wrong because CloudTrail records who called which API and when, not the full prompt and response text.
• B is wrong because an in-memory cache is temporary and can be changed.
• C is wrong because metrics hold counts and numbers, not the content of prompts and responses.

Prompt caching stores a repeated prompt prefix so later requests read it from cache instead of processing it again. That lowers the cost of the cached input tokens and reduces latency for the 10–20 follow-up questions.

• A is wrong because Provisioned Throughput reserves capacity; it does not reduce the cost of resending the same tokens.
• C is wrong because batch inference is asynchronous and unsuitable for a user asking questions interactively.
• D is wrong because output length does not address the repeated input tokens.

Batch inference processes large sets of prompts asynchronously from S3 and writes results back to S3, at a lower price than on-demand calls for supported models. A 48-hour window suits it.

• B is wrong because on-demand calls at full speed cost more and will hit throttling limits.
• C is wrong because paying for a month of reserved capacity for a once-a-month job wastes money.
• D is wrong because streaming improves perceived latency for a waiting user, and nobody is waiting here.

The default parser only outputs text, so tables lose their structure. Bedrock Data Automation and foundation model parsers are built to extract tables, figures and charts from PDFs, which fixes the chunks at the source.

• A is wrong because more jumbled chunks are still jumbled.
• B is wrong because a larger model cannot recover table structure that was lost during ingestion.
• C is wrong because a higher temperature makes the model guess, which increases wrong torque values.