Free AIP-C01 practice test: 30 questions with answers

Knowledge base content is indexed when an ingestion (sync) job runs, so edits in S3 are not reflected in the vector index until the data source is re-synced. Changing the model or raising the retrieval count does not fix stale indexed content, and model access grants are unrelated to indexing.

Similarity search only works when queries and documents are embedded by the same model, because each model defines its own vector space. Chunk size, reranking, and the index algorithm may affect quality, but mismatched embedding models make the distances meaningless.

Contextual grounding checks score a response for grounding in the supplied source text and for relevance to the query, and they can block responses that fall below the thresholds. Word filters match literal terms, while Provisioned Throughput and invocation logging do not evaluate content.

Metadata filters narrow the candidate set to documents that are valid for the user and the question, and hybrid search improves recall for exact terms such as product codes that pure vector search can miss. Output limits and generator temperature do not affect which chunks are retrieved.

Overlap copies text near each boundary into the neighbouring chunk, so a clause that crosses a split can still be retrieved intact. Indexing whole PDFs usually dilutes the embedding and hurts retrieval precision, and reducing dimensions or raising top-k does not fix the split.

Amazon Macie uses machine learning and pattern matching to discover and classify sensitive data such as PII in S3 buckets. Inspector scans workloads for vulnerabilities, GuardDuty detects threats in activity logs, and AWS Config records resource configuration.

An action group describes the callable operations with an OpenAPI schema, and the agent invokes a Lambda function or API integration to perform them. A knowledge base is for retrieval over documents, and prompt text cannot execute an API call.

InvokeModelWithResponseStream returns the generated output as a stream of chunks, which lets the UI render partial text. InvokeModel returns the complete response at once, batch jobs are asynchronous, and Retrieve only returns knowledge base passages.

Step Functions coordinates multi-step workflows with built-in retries, error handling, parallel Map states, and callback patterns for human approval. SNS and SQS move messages but do not manage workflow state, and Route 53 is DNS.

Validating inputs before writes stops malformed or unsafe actions, and structured error messages let the agent recover or ask for help. Broad administrator access and default access to every API widen the blast radius, and removing timeouts can hang the agent.

EventBridge routes the S3 event to targets through rules, which decouples producers and consumers and avoids polling. Scheduled or manual batches and polling add latency and operational work.

Multi-agent collaboration lets a supervisor agent route subtasks to collaborator agents and merge the results. The other options control evaluation, capacity, or cost of a single model call.

Guardrail sensitive information filters detect PII in prompts and responses and can block or mask it at invocation time. Invocation logging records traffic but does not filter it, while throughput and batch settings do not inspect content.

Clear separation of untrusted content, least-privilege tool access, and output validation limit what a successful injection can do. Models cannot be relied on to ignore instructions in their input, and placing user content in the system prompt makes it more trusted, not less.

An AWS Organizations service control policy sets a permission ceiling across accounts, so only approved model ARNs can be invoked. Temperature, documentation, and tags do not restrict access.

Pre-launch evaluation on representative inputs surfaces harmful or biased behaviour, and clear AI disclosure with a human escalation path supports transparency and recourse. Hiding AI use, skipping monitoring, and removing feedback each weaken oversight.

A lifecycle expiration rule deletes objects automatically after the configured age. Versioning keeps old copies, Object Lock prevents deletion for its retention period, and transitioning to archive storage keeps the data instead of removing it.

High-risk outputs should pass through human review, carry clear disclaimers, and be checked against sources where possible. Automatic publishing, removing citations, and higher temperature each increase the risk of confident errors.

Model routing sends easy requests to a cheaper model and keeps the expensive model for hard cases, which lowers average cost. Larger outputs and removing caching increase cost, and consistency does not require the largest model everywhere.

Prompt caching stores the processed prefix of a prompt so repeated requests reuse it, reducing cost and time on supported models. Batch inference handles offline jobs, Guardrails filter content, and knowledge base sync updates the index.

Provisioned Throughput reserves model capacity for consistent latency and throughput, which suits steady production traffic. On-demand capacity can be throttled under load, and batch inference is asynchronous rather than real-time.

Sending fewer but more relevant chunks cuts input tokens while keeping the evidence, and concise instructions with an output limit cap token use. Sending the whole knowledge base or long answers and repeating text all increase tokens without improving the answer.

CloudWatch publishes Bedrock runtime metrics such as invocation latency, token counts, and throttles, and supports dashboards and alarms. Macie finds sensitive data, Athena queries stored data, and Backup protects data copies.

Target-tracking autoscaling adds and removes instances based on a utilization metric, keeping cost close to demand while meeting latency targets. Fixed fleets pay for idle peak capacity, and manual scaling responds too slowly.

A golden dataset with automated metrics gives repeatable comparisons against a baseline, which makes regressions visible before release. Single spot checks, waiting for complaints, and self-reported confidence are not reliable signals.

Faithfulness checks that the summary does not add unsupported claims, and coverage checks that important points are retained. Training hardware and release date say nothing about the quality of a specific output.

Agent traces show which action was chosen and why, so the team can fix unclear descriptions or schemas at the source. Timeouts, disabling the agent, and chunk size do not address action selection.

A labelled sample of real tickets with a rubric measures quality on the actual task, and human confirmation checks the automated grading. Leaderboards, parameter counts, and answer length do not measure how well a model performs this task.

Per-stage timing shows whether the delay is in embedding, retrieval, reranking, or generation, and index settings such as top-k directly affect search cost. The frontend, bucket name, and browser cache are unrelated to retrieval latency.

A measured quality regression should block promotion until the cause is understood, so the team holds or reverts the change and examines failing cases. Ignoring or hiding the metric and shipping regardless bypass the control the evaluation exists to provide.