Bedrock Responsible AI Design Lab

Goal

Reinforce Prompt risk, RAG grounding, Guardrails, Human review, Data handling through a small, inspectable activity. This is a learning exercise, not a production design: use a dedicated sandbox or test account, record what you observe, and do not copy the permissions or data-handling choices into production without review.

What you should be able to explain afterward: what AWS managed for you, what you configured, what evidence proves the result, and what could continue to cost money or expose data if you leave it behind.

Concepts Reinforced

• Prompt risk
• RAG grounding
• Guardrails
• Human review
• Data handling

Prerequisites

• No AWS account required for design-only path

Before you start

1. Choose one AWS Region and write it down; many resources and console views are Region-specific.
2. Confirm that you are operating only in an authorized non-production account and that you can identify the account ID before creating anything.
3. Open the AWS service documentation linked below if a console label or command option is unfamiliar. The lab is successful when you understand the observation, not when you click through it quickly.

Estimated Cost

Exact Steps

Read the entire sequence before beginning. Substitute your own test names for every placeholder, keep the Region consistent, and pause after each step to inspect the result. If a command returns an error, do not repeatedly retry it without reading the error: check Region, account, permissions, resource identifiers, and whether the preceding step actually completed.

1. Choose a realistic business assistant scenario.
2. Classify prompt and response data sensitivity.
3. Decide whether the assistant needs public model knowledge, internal documents, or both.
4. Add guardrail categories, logging controls, and human-review triggers.
5. Write acceptance tests for hallucination, unsafe output, and source-grounding failures.

Block: regulated advice without reviewer | Mask: secrets and identifiers | Log: prompt metadata only | Review: low confidence or high impact

Verification

Verification is the evidence that the concept worked. Capture the relevant ARN, status, identity, log entry, policy result, object version, or query output before cleanup. A successful command alone is not proof that the intended control behaved correctly.

• Sensitive data has an access and retention decision.
• A high-impact answer path includes human review.
• Grounded answers cite approved sources in the design.

If the result is different

• Confirm the selected Region and AWS account.
• Check the exact resource identifier and current status.
• Review the service event history, CloudTrail event, CloudWatch log, or command output when available.
• Re-read the relevant IAM policy, trust policy, security rule, or service setting instead of assuming the service is at fault.

Cleanup Steps

Cleanup is part of the lab. Remove test resources in dependency order, delete temporary credentials or local files, and check the billing or resource console for anything that remains. Some services retain versions, snapshots, logs, or recovery artifacts even after the visible parent resource is deleted.

• No cleanup for design-only path. Delete any test knowledge base, vector store, logs, or model invocation artifacts if created.

Sources and Review Metadata