Multi-Tenant RAG with Amazon Bedrock: Secure SaaS AI

This article details a secure, multi-tenant Retrieval Augmented Generation (RAG) implementation using Amazon Bedrock and OpenSearch Service. The solution addresses the challenges of providing personalized AI services to multiple SaaS tenants while maintaining strict data isolation. A key feature is the use of JSON Web Tokens (JWT) combined with OpenSearch's Fine-Grained Access Control (FGAC) to manage access permissions dynamically and scalably. This avoids the limitations of managing numerous IAM roles for each tenant. The architecture leverages several AWS services: Amazon Cognito for user authentication and JWT generation, DynamoDB for tenant metadata and routing information, Amazon S3 for document storage, and AWS Lambda for orchestration. The system offers three data isolation patterns: domain-level, index-level, and document-level, allowing for flexibility based on security needs. Amazon Bedrock‘s language models handle vectorization and LLM processing. The solution's advantages include dynamic tenant identification, seamless FGAC integration, and scalability. However, it requires using OpenSearch Service, not OpenSearch Serverless, due to limitations in the latter's permission model. The article also notes that for production, shared DynamoDB and S3 resources might need partitioning for optimal performance and cost efficiency. The target audience is SaaS providers looking to build personalized AI capabilities into their applications with strong security and scalability. The provided implementation uses Anthropic's Claude 3.5 Sonnet v2 and Amazon Titan Text Embedding V2, but other models could be substituted. While the solution offers robust security and flexibility, the complexity of setting up the multi-service architecture should be considered.

(Source: https://aws.amazon.com/blogs/machine-learning/multi-tenant-rag-implementation-with-amazon-bedrock-and-amazon-opensearch-service-for-saas-using-jwt/)