When businesses explore RAG systems, security and data sovereignty are top concerns. Where does your data go? Who can access it? Does it leave the country? These questions matter—especially for businesses in regulated industries or those handling sensitive client information.
AWS provides the infrastructure to build RAG systems where your data stays completely under your control, in Australian data centres, protected by enterprise-grade security. Here's how it works.
The Architecture
A secure RAG system on AWS typically uses these components:
Amazon S3 — Document Storage
Your source documents (PDFs, manuals, policies) are stored in S3 buckets in the ap-southeast-2 (Sydney) region. S3 provides encryption at rest (AES-256), versioning, and fine-grained access controls. Documents never leave the Sydney region.
Amazon OpenSearch — Vector Database
OpenSearch stores the embeddings (vector representations) of your document chunks. It provides fast similarity search across millions of vectors. Running in a private VPC means it's not accessible from the public internet.
AWS Lambda — Processing
Serverless functions handle document ingestion, embedding generation, query processing, and response generation. You only pay for what you use, and there's no server infrastructure to maintain or secure.
Amazon Bedrock — AI Models
AWS Bedrock provides access to foundation models (Claude, Titan, etc.) that run within AWS infrastructure. Your data stays within AWS's security boundary—it's not sent to third-party AI providers.
VPC — Network Isolation
All components run inside a Virtual Private Cloud. Private subnets, security groups, and network ACLs ensure that your RAG system isn't exposed to the public internet.
Security Controls
Enterprise Security Layers
- IAM (Identity & Access Management). Role-based access ensures only authorised users and services can access specific resources. Principle of least privilege applied throughout.
- Encryption at rest. All data encrypted using AWS KMS (Key Management Service) with customer-managed keys.
- Encryption in transit. All communication between components uses TLS 1.3.
- VPC endpoints. Internal AWS services communicate privately without data traversing the public internet.
- CloudTrail logging. Every API call is logged for audit purposes. Who accessed what, when, and from where.
- WAF (Web Application Firewall). Protects the user-facing API from common web attacks.
Australian Data Sovereignty
For businesses subject to Australian Privacy Principles or industry-specific regulations, data sovereignty is non-negotiable. Our AWS RAG architecture ensures:
- All data in ap-southeast-2. Documents, embeddings, logs, and backups all reside in the Sydney region.
- No cross-region replication. Data is not replicated to overseas regions unless explicitly configured.
- AWS compliance. AWS Sydney meets IRAP (Information Security Registered Assessors Program) requirements.
- Contractual protections. AWS Enterprise agreements can include data residency commitments.
Critical for: Medical clinics handling patient data, law firms with client privilege obligations, financial advisers with ASIC compliance requirements, and government contractors with security clearance needs.
Document-Level Access Control
Not everyone in your organisation should have access to all documents. A properly designed RAG system implements:
- Role-based access. HR documents visible to HR team, financial documents to finance team.
- Matter-level isolation. For law firms—Client A's documents never appear in Client B's queries.
- Classification levels. Public, internal, confidential, restricted—each tier with appropriate access controls.
- Attribute-based access. Access based on user attributes like department, location, or seniority level.
Typical Cost Structure
Running a secure RAG system on AWS is more affordable than many businesses expect:
| Component | Typical Monthly Cost |
|---|---|
| S3 storage (documents) | $5-50 depending on volume |
| OpenSearch (vector DB) | $200-800 depending on size |
| Lambda (processing) | $20-100 depending on queries |
| Bedrock (AI model usage) | $50-500 depending on volume |
| Other (VPC, CloudWatch, etc.) | $30-100 |
| Total | $300-1,500/month typical |
Compare this to the cost of senior staff time spent answering repetitive questions, and the ROI becomes clear. See our Calculating AI ROI guide for a detailed framework.
Getting Started
- Define your security requirements. What regulations apply? What data classifications do you have?
- Identify your first use case. Start with a bounded document set and user group.
- Design the architecture. Map components to your security requirements.
- Build incrementally. Start with basic RAG, then add security layers and access controls.
- Test with real users. Verify both functionality and security before production deployment.
HELLO PEOPLE specialises in building secure, AWS-hosted RAG systems for Australian businesses. We handle the architecture, security, and deployment so you can focus on the business outcomes. Talk to our AI team.
