RAG data boundary lab

Goal: learn retrieval-augmented generation by focusing on the security question first: what is the model allowed to retrieve?

Outcome

You should finish with a simple RAG design that separates public learning docs from private or restricted data, then documents how access is controlled.

Example AWS shape

  • Data source: S3 bucket with public-safe sample docs only.
  • Indexing path: Amazon Bedrock Knowledge Bases or a small vector-search lab.
  • Encryption: KMS-managed encryption for storage where applicable.
  • Access: IAM role scoped to only the approved bucket or collection.
  • Logging: CloudTrail for API activity and CloudWatch for app behavior.

Runbook steps

  1. Create a public-safe sample document set. Do not use confidential data.
  2. Label each document as public, internal-test, or restricted. Only public-safe docs go into the lab.
  3. Write the retrieval boundary before building the index.
  4. Build or diagram the retrieval flow.
  5. Test questions the system should answer and questions it should refuse or not know.
  6. Record where logs and access records would be reviewed.

Questions to answer

  • Can the model access only the intended documents?
  • What happens if a user asks for data outside the allowed set?
  • Where could sensitive information leak into logs?
  • Who can update the source documents?
  • How would stale or poisoned documents be removed?

Portfolio artifact

Create a RAG boundary diagram plus a short test plan. Show what data is allowed, what data is blocked, who can update it, and how retrieval behavior is checked.