AIP-C01 Timed Hard Questions — 25 Difficult Questions

Question 1

A healthcare company is building a generative AI assistant on Amazon Bedrock that summarizes patient records. The compliance team requires that the assistant's outputs are traceable and that any harmful or biased content is detected before being shown to clinicians. The company wants to use AWS-native services and features. (Choose two.)

15 minute time limit — choose an answer to begin.

Accepted Answer

Enable model invocation logging in Amazon Bedrock to capture request and response metadata to Amazon CloudWatch Logs and Amazon S3.. Amazon Bedrock Guardrails evaluates prompts and completions against configurable policies to block harmful or biased content before it is presented, while model invocation logging captures detailed request and response data for traceability. Together they provide both preventive content moderation and a durable audit trail. AWS WAF, Amazon Macie, and AWS Trusted Advisor operate at different layers and cannot inspect or record natural language content for bias or harmful material.

Answer

Use AWS Trusted Advisor to review the Bedrock model for ethical AI compliance.

Answer

Enable Amazon Macie to automatically redact biased statements from Bedrock responses.

Answer

Configure AWS WAF to inspect the generative AI responses for biased language.

Question 2

You are implementing a multi-agent system. Which method is most effective for ensuring the agents do not loop infinitely when coordinating tasks?

Accepted Answer

Implementing a maximum iteration count. A maximum iteration count (or recursion/step limit) is the standard guardrail in agent orchestration frameworks like LangChain, LangGraph, and AWS Bedrock Agents to bound execution. When agents can call each other or themselves, a hard cap on turns prevents runaway loops that burn tokens and never terminate. Once the cap is hit, the orchestrator returns the best partial result or raises a controlled error.

Answer

Caching agent responses

Answer

Increasing the model temperature

Answer

Using a smaller model

Question 3

Which THREE factors influence the cost of running an Amazon Bedrock Knowledge Base?

Accepted Answer

The amount of data stored in the vector database.. Option A is correct because an Amazon Bedrock Knowledge Base stores its embeddings in a vector store (for example, Amazon OpenSearch Serverless, Aurora pgvector, or Neptune Analytics), and the cost of that vector database scales with the volume of data/vectors stored. Option B is correct because the RetrieveAndGenerate API invokes a foundation model to generate the final answer, and you are billed for the input and output tokens the model consumes during that generation step. Option D is correct because ingesting documents into the knowledge base requires calling a Bedrock embedding model (such as Amazon Titan Text Embeddings), and you pay per token processed by that embedding model. Option C is not a Bedrock Knowledge Base cost factor because IAM users are free and concurrency of users does not drive Bedrock or vector-store charges. Option E is not correct because there is no per-scan charge for S3 buckets; S3 charges are based on storage, requests, and data transfer, not on how often the knowledge base syncs or scans the bucket.

Answer

The number of concurrent IAM users.

Answer

The frequency of S3 bucket scans.

Question 4

How can a developer ensure that the retrieval step in a RAG application focuses only on the most relevant documents when the knowledge base contains millions of files?

Accepted Answer

Implement metadata filtering in the query.. Metadata filtering (A) allows the retrieval step to narrow down the search space by applying filters on document attributes (e.g., date, category, author) before performing vector similarity search. This ensures that only documents matching the filter criteria are considered, improving relevance and reducing noise when the knowledge base is large.

Answer

Increase the vector dimension size.

Answer

Use a larger model for embedding.

Answer

Reduce the chunk size.

Question 5

A developer is designing a RAG system and wants to ensure high-quality retrieval. Which THREE techniques should be implemented to improve semantic search relevance?

Accepted Answer

Using a reranking model to score retrieved documents. Option A is correct because a reranking model (e.g., a cross-encoder) re-scores the initially retrieved candidates against the query, capturing fine-grained query-document interactions that bi-encoder vector similarity misses, which directly boosts top-k relevance. Option B is correct because semantic chunking based on document structure (headings, sections, paragraphs) preserves coherent meaning units, so each embedding represents a complete idea rather than an arbitrary text slice, improving retrieval precision. Option D is correct because hybrid search combines lexical keyword matching (e.g., BM25) with dense vector search, capturing exact terms, rare entities, and acronyms that pure embeddings often miss while retaining semantic matching. Option C is not correct because simply maximizing vector dimensions increases storage and compute cost without guaranteeing better relevance, and can even hurt performance through overfitting or noise. Option E is not correct because fixed-size overlap chunking ignores semantic boundaries, frequently splitting related content and diluting embedding quality compared to structure-aware chunking.

Answer

Increasing the vector dimension size to maximum

Answer

Using only fixed-size overlap chunking

Question 6

When migrating an on-premises RAG system to Amazon Bedrock, how should the team handle existing vector embeddings?

Accepted Answer

Re-embed the documents using a model supported by Amazon Bedrock.. When migrating a RAG system to Amazon Bedrock, existing vector embeddings from an on-premises system cannot be directly used because they were generated by a different embedding model. Embeddings are model-specific; dimensions and semantic spaces differ. Therefore, the documents must be re-embedded using an embedding model supported by Amazon Bedrock (e.g., Amazon Titan Embeddings or Cohere Embed). This ensures compatibility with Bedrock's vector store and retrieval.

Answer

Map the old embeddings to new dimensions using a linear function.

Answer

Use the Bedrock Import Vector API.

Answer

Import existing vectors directly into the OpenSearch index.

Question 7

You are using SageMaker JumpStart and need to enable 'Data Capture' for monitoring model inputs and outputs. Where is the captured data stored?

Accepted Answer

Amazon S3. When you enable Data Capture for a SageMaker endpoint (including JumpStart models), the captured data (inputs and outputs) is stored in an Amazon S3 bucket that you specify. This allows you to later analyze the data for monitoring, debugging, or retraining. CloudWatch Logs is used for logging but not for storing the captured data itself; the data capture feature writes to S3.

Answer

Amazon RDS

Answer

Amazon CloudWatch Logs

Answer

Amazon ElastiCache

Question 8

You are building an agentic application. Which THREE of the following are valid ways for an agent to perform an action?

Accepted Answer

Invoking a Lambda function. Option B is correct because invoking a Lambda function is a canonical way for an agent to execute an action through a defined tool/function interface, letting the agent trigger server-side compute with specific parameters. Option C is correct because querying a Knowledge Base is a valid agent action that retrieves relevant information (typically via retrieval APIs such as Amazon Bedrock Knowledge Bases Retrieve/RetrieveAndGenerate) to ground the agent's next step. Option E is correct because returning a direct response to the user is itself a valid action the agent can take, completing the turn with a natural-language answer when no external tool call is needed. Option A does not belong because directly modifying a database table without an API bypasses controlled, auditable tool interfaces and is not a supported agent action pattern. Option D does not belong because running a local Docker container is an infrastructure/runtime operation, not an agent action exposed through the agent's tool/action model.

Answer

Directly modifying a database table without an API

Answer

Running a local Docker container

Question 9

An enterprise requires that all Generative AI model outputs are logged for compliance, but logs must not contain any PII. How can this be achieved in an AWS environment using Amazon Bedrock?

Accepted Answer

Enable Bedrock Model Invocation Logging with Guardrails PII redaction enabled. Bedrock Model Invocation Logging captures full request and response payloads to CloudWatch Logs or S3, and when combined with Bedrock Guardrails configured with PII detection and redaction, sensitive entities are masked before the log is written. This satisfies both the compliance requirement (full audit trail of model interactions) and the privacy requirement (no PII in logs) without custom code. Guardrails supports built-in PII filters for categories like NAME, EMAIL, SSN, and PHONE.

Answer

Write a Lambda function to strip logs before writing to S3

Answer

Use IAM policies to prevent logs from being written

Answer

Use AWS CloudTrail to log API calls and enable redaction in CloudWatch

Question 10

A researcher wants to ensure the model output does not reference external websites. How can this be enforced?

Accepted Answer

Configure a Guardrail with a Denied Topics policy. Amazon Bedrock Guardrails provide a Denied Topics policy that can block model outputs referencing specified topics, including external websites or URLs. Configuring a guardrail with a denied topic for external references enforces the restriction at the service level, independent of prompt wording. This is the correct mechanism for enforcing output constraints.

Answer

Use a custom system prompt to tell the model not to use URLs

Answer

Disable model streaming

Answer

Use IAM to block internet access for the application

Question 11

A developer is building a RAG system and wants to include document metadata in the search retrieval. How is this done?

Accepted Answer

By adding metadata files alongside the documents in S3 and using custom ingestion logic.. Metadata is attached to the chunks in the vector index. During query, the system can use the `filter` parameter to limit results.

Answer

By prepending the metadata to the prompt.

Answer

It is done automatically by the model.

Answer

It is not possible to use metadata in Bedrock RAG.

Question 12

Which configuration in an Amazon Bedrock Knowledge Base is essential for updating the vector index whenever the source files in S3 change?

Accepted Answer

Configure Data Source Syncing.. Syncing is triggered via the 'Sync' operation, but for automation, S3 Event Notifications coupled with a Lambda function or the Bedrock Sync API is required.

Answer

Use AWS Config to monitor bucket changes.

Answer

Set the knowledge base to 'Auto-Refresh' mode.

Answer

Enable S3 Versioning.

Question 13

Your Generative AI application uses a vector database on Amazon OpenSearch Service. You notice high CPU usage during ingestion. How can you optimize performance?

Accepted Answer

Use bulk requests with an optimized refresh interval. Increasing the number of dedicated primary shards or using a more performant instance type for the ingestion nodes are standard optimization patterns.

Answer

Reduce the index replication factor

Answer

Disable vector search

Answer

Increase the number of replica shards

Question 14

An organization is integrating a custom model via Bedrock Custom Models. They need to ensure that the data used for fine-tuning remains within their VPC. Which configuration meets this requirement?

Accepted Answer

Configuring an interface VPC endpoint for Amazon Bedrock. AWS PrivateLink for Amazon Bedrock allows the traffic to never traverse the public internet, keeping the data flow contained within the VPC context.

Answer

Using AWS Glue for data preparation only

Answer

Deploying the model in an on-premises data center

Answer

Encrypting the dataset with AWS CloudHSM

Question 15

Which THREE factors influence the choice of embedding model for a RAG solution?

Accepted Answer

Supported languages. Language support, embedding dimensions, and maximum input token length are key factors.

Answer

Model's popularity

Answer

Model training cost

Question 16

A developer needs to validate that a fine-tuned model does not exhibit toxic behavior. Which AWS service should be used to automate this validation?

Accepted Answer

Amazon Bedrock Evaluation jobs with built-in toxicity metrics.. Amazon Bedrock Evaluation jobs allow for built-in toxicity and bias testing against standard benchmarks.

Answer

Amazon SageMaker Model Monitor.

Answer

A

Answer

AWS Glue to sanitize data.

Answer

AWS Lambda to call the model and manually inspect outputs.

Question 17

Which THREE techniques can improve the performance (latency) of an inference endpoint?

Accepted Answer

Model quantization. Model optimization, hardware-specific instance types, and batching are standard techniques for improving inference speed.

Answer

Moving model artifacts to S3 Standard

Answer

Increasing the number of replicas

Question 18

A company is using Amazon Bedrock with Guardrails to filter PII. Users report that some PII is still passing through. What is the best approach to troubleshoot this?

Accepted Answer

Enable the 'trace' functionality in the InvokeModel API response to analyze Guardrail findings.. Checking the Guardrails assessment trace via the InvokeModel API response is necessary to see if the filter was triggered or bypassed.

Answer

A

Answer

Disable the Guardrail and re-run the test to see if performance improves.

Answer

Route all traffic through a WAF to block PII patterns.

Answer

Increase the PII sensitivity threshold in the IAM policy.

Question 19

When troubleshooting a SageMaker endpoint that consistently returns 503 Service Unavailable errors, which THREE areas should you investigate?

Accepted Answer

Instance memory usage. 503s typically point to capacity limits, health check failures, or resource exhaustion (memory/GPU).

Answer

IAM role permissions

Answer

Input JSON schema validation

Question 20

You are configuring a Guardrail for a multilingual application. What is the best practice for configuring content filters?

Accepted Answer

Create one Guardrail per language. Configure multiple language-specific filters or ensure the Guardrail content filter is configured to handle the detected language of the input.

Answer

Disable filters for non-English languages

Answer

Translate all inputs to English first

Answer

Use a regex-based filter only