AWS rethinks access control for RAG: how Amazon Quick and Bedrock protect corporate bases

RAG access control: The problem AWS wants to solve

AWS has published an approach that combines Amazon Quick and Amazon Bedrock to solve a silent neck of corporate RAG: ensuring that responses generated by AI respect original access permissions from sources such as SharePoint, Google Drive and Confluence. In practice, the model should only recover and synthesize excerpts that the user would already have the right to see directly in the source — no lateral leaks between teams, areas or contracts.

According to the official machine learning blog of AWS(https://aws.amazon.com/blogs/machine-learning/rethinking-access-control-for-rag-with-amazon-quick-and-amazon-bedrock), the starting point is that filtering permissions cannot be treated as implementation detail. She needs to be at the center of the recovery architecture design.

Why does traditional RAG fail access control in regulated environments?

The classical flow of RAG indexes documents, generates embeddings and delivers the most semantically close excerpts of the question. The problem is, semantic similarity doesn't know who's asking. A junior analyst and a legal director can do the same query and receive the same confidential contract excerpt, even if only one of them is allowed to read it.

In regulated sectors — legal, health, finance, public sector — this misalignment is not just a technical bug. It is a data leak vector, with direct implications in LGPD, confidentiality contracts and internal access segregation policies. This is what AWS' approach aims to address by tying Amazon Quick to Amazon Bedrock: Quick acts as a layer of experience and governance for corporate knowledge bases, while Bedrock provides model infrastructure and recovery.

What changes in RAG access control with Amazon Quick and Bedrock?

The central idea is to inherit permissions directly from the original sources, rather than manually replicate them in a proper index. This reduces the risk of divergence between who can access what in SharePoint and what the AI assistant delivers.

For engineering teams, the practical impact is relevant: less time spent building a proprietary layer of ACLs (Access Control Lists) and more focus on recovery quality and responses. It is a line close to other AWS initiatives in this direction, such as sandboxes that restrict what agents read and modify, a theme already covered in the portal in AWS launches Strands Box: sandboxes that control what AI agents read and modify.

RAG access control: what changes to Brazilian companies?

Brazilian organizations that already use Google Workspace, Microsoft 365 or Confluence as central repositories find here a more direct way to get RAG out of the lab and put it into production without giving up governance. Sectors such as banks, insurance companies, legal offices and healthtechs — the most exposed to LGPD — tend to be the first to require such collateral before releasing internal assistants.

It is worth remembering that safe adoption also requires governance of people: defining roles, reviewing orphan permissions and maintaining continuous auditing. The tool solves the technical layer, but does not replace well-designed access policy.

RAG Access Control: What AWS Approach Still Not Solve

The AWS publication describes the design of the solution, but does not enter public latency benchmarks, cost per thousand queries or numerical comparison with market alternatives. In the material there is also no closed list of connectors approved other than those already mentioned. For teams operating on a scale, these numbers will only come from proofs of their own concept — and that is precisely where the access control architecture needs to be validated before it becomes critical dependence.

Frequently Asked Questions about RAG access control with AWS

What exactly does AWS propose with Amazon Quick and Bedrock for RAG?

It proposes an architecture that combines Amazon Quick as a layer of experience and governance and Amazon Bedrock as a model engine, ensuring that AI responses respect permissions from original sources such as SharePoint, Google Drive and Confluence.

Why is this critical in regulated environments?

Because the traditional RAG filters by semantic similarity, not by user permission — which can expose confidential snippets to those who would not have direct access to the source, creating risk of leakage and non-compliance with LGPD.

Can Brazilian companies already use this approach in production?

AWS published the solution design, but cost, latency and complete list of connectors were not detailed. Brazilian Times need to validate architecture in concept tests before treating it as critical dependence.