Skip to content
Artificial intelligenceOctober 12, 20263 min · updated October 15, 2026

AWS documents real-time RAG access controls for business documents

Daillac research
technical feedback from AWS
D
Written by
Daillac

Editorial team

Document passing through two control points where the second denies access
On October 7, 2026, AWS published an access control architecture for RAG document search with Amazon Quick and Amazon Bedrock. It combines filtering of permissions stored in the index and verification with the source at the time of the request. The case to monitor is that of a document still indexed after an access right has been revoked.

An indexed permission can become outdated

The technical feedback from AWS describes two control steps. The system selects documents according to the indexed permissions, and then checks the current permissions with the document source. This second check is aimed at access changes that have occurred since indexing.
In a fictitious example, an employee leaves a project and loses access to their document folder. The search engine still retains an indexed copy of a report. A query made to the assistant must respect the removal of the right, even if the document matches the question very well.
This situation also concerns the excerpts that the assistant sends to the model. The check must occur before the refused content enters the response context. Hiding the citation afterward would leave the model having received information it should not have used.

Testing access removal

Create a test document and allow two accounts to view it. Index it, then remove access for one of the accounts in the source. Ask the same question with each. The authorized account must be able to retrieve the content according to your rules; the other must receive a response that does not reveal the document.
Repeat the test with an already open conversation. Check if contents read before the removal remain in session memory and how your policy handles this case. The results cache, summaries, and logs must be included in this review.
Also define behavior when the source does not respond. For a sensitive document, an impossible verification may need to block its use. This rule must be decided with the data manager and applied in the search system.

Verify permissions over the entire process

A document assistant can combine sources with different rights. Examine the connector accounts, synchronized groups, and update frequency. A chat application’s permissions do not necessarily describe the permissions of each document.
The MCP vs API and tool controls guide details the role of the execution system. For literature search, keep track of the access decision with the minimum content needed for diagnosis.
The RAG and knowledge base service allows you to scope a pilot with the sources, rights, and retirement cases to be tested. The AWS release presents an architecture; the connectors and conditions of your deployment must be verified before adopting it.

Review the permissions of your document database

Present your needs and the relevant tools to scope the useful checks for your project.
Sources & method

Official AWS publication of 2026-10-07, accessed on October 12, 2026. Summary of documented facts and Daillac analysis; explicitly fictional examples, without vendor tests or attributed customer results.

Share