AIP-C01 sample exam questions

AWS AppConfig delivers routing rules and model choices that can change without a deployment, and the Lambda agent caches them for fast reads across many requests. The function then applies the complex business logic per request. Environment variables and stage variables need updates and cannot express these rules, and authorizers are meant for access decisions.

Bedrock Knowledge Bases with GraphRAG on Neptune Analytics automatically builds relationships between entities and uses them during retrieval, which captures multi-hop connections with little management. Vector search alone finds similarity rather than connected relationships, and the other options need custom relationship logic.

AWS Glue Data Catalog registers data sources and supports tags that attribute content to them, and CloudTrail logs access across services. CloudWatch Logs does not audit access by itself, while S3 tags and Lake Formation cover only part of the requirement for internal and external sources.

A Step Functions Parallel state runs the analyses at the same time, so total time falls toward the longest branch rather than the sum, and CloudWatch shows timing. The other options still process work sequentially. An SQS queue also adds wait time, and CPU-based ECS scaling does not help calls that wait on a model.

The Step Functions Bedrock integration can read input from and write output to S3 URIs, so states pass small references rather than large payloads. ResultPath and ResultSelector route those references using the existing bucket, and the sequential workflow stays visible. Compression only delays the limit, and DynamoDB or extra state machines add work.

A version-controlled repository keeps prompt history, and a test suite with complex documents and quantifiable metrics shows where and why results change. Testing only simple documents misses the problem, splitting production traffic risks clinical output, and Comprehend Medical extracts entities rather than scoring factual accuracy.

Console-based Bedrock evaluation jobs read datasets from S3 and require IAM access plus CORS on the bucket. Built-in jobs measure real world knowledge for text generation and BERT Score for summarization, which are accuracy metrics. VPC endpoints are not the requirement here, and toxicity or word error rate are not accuracy measures for these tasks.

A semantic cache compares embeddings, so differently worded questions can match a stored answer. MemoryDB performs this search in memory with very low latency, and a similarity range finds close matches. DAX and DynamoDB match text rather than meaning, and OpenSearch can do semantic matching but is slower for caching.

Bedrock Guardrails can detect and filter PII and deny investment advice topics with no custom code, and Bedrock model invocation logging can deliver text and images to S3. Macie scans S3 rather than live prompts, Comprehend topic modeling is not a real-time control, and regex with prompt instructions is not reliable.

The Converse API offers one consistent request format through both the Python and JavaScript SDKs, and IAM roles work in Lambda and EKS. Multi-turn context is passed in the messages array. InvokeModel needs model-specific formats, ElastiCache adds a service that is not needed, and process memory would be lost between runs.

Bedrock Intelligent Prompt Routing automatically sends each request to a suitable model in the same family based on predicted quality and cost, with no routing code. A classifier stage with Lambda takes more effort and adds latency, one mid-sized model is not cost-optimized, and keyword rules are fragile.

A reference dataset of validated question-answer pairs from clinical guidelines gives a trusted comparison, and Bedrock evaluation measures factual accuracy before release, including across rephrased questions. Keyword checks and pattern rules cannot detect subtle fabrication, and alarms alone do not measure accuracy. The output diffing with Guardrails wording in B is loose, but this is the closest fit.

An AgentCore outbound credential provider can obtain temporary credentials through AssumeRole, one role per account gives clear security boundaries, and session tags record the end user in CloudTrail. Adding accounts only needs another role. A separate assistant per account is much more work, DynamoDB tables are not shared through AWS RAM here, and SCPs do not grant access.

A European geographic inference profile keeps European data in European Regions while still spreading load across them, and on-demand inference needs no capacity management. Route 53 health checks handle failover, and CloudWatch monitors utilization. Provisioned throughput per Region requires custom routing, Bedrock provisioned capacity lacks standard automatic scaling, and streaming is not needed.

The service role's permissions are in place, but Bedrock also needs a trust policy naming it as the service principal so it can assume the role, and its absence causes AccessDenied. CORS applies to browser access, and the dataset or metric choices do not affect access. TREX and Accuracy are valid for general text generation.