Build multi-tenant agentic chat applications on enterprise data with Amazon Bedrock Managed Knowledge Base

This article describes how to build multi-tenant agentic document chat applications using Amazon Bedrock Managed Knowledge Base with secure per-user data isolation and asynchronous ingestion.

• Amazon Bedrock Managed Knowledge Base handles ingestion, storage, embedding, and ranking without infrastructure provisioning
• Agentic retrieval decomposes complex questions into sub-queries with multi-hop retrievals while enforcing per-user isolation on every hop
• Direct ingestion via custom connector allows documents to become retrievable within seconds; files up to 6 MB sent inline, larger files staged to S3
• Documents progress through five indexing states (STARTING, PENDING, IN_PROGRESS, TEXT_INDEXED, INDEXED) with text queryable in 2-30 seconds
• Per-user isolation enforced through metadata filtering on user_id derived from verified JWT, not client-supplied values
• Decouple uploads from ingestion using SQS queue; batch up to 10 documents per API call for throughput optimization
• Monitor ingestion success rate and time-to-INDEXED at P50/P99 as early indicators of throughput ceiling
• Store conversation history in DynamoDB and pass relevant turns with each request to AgenticRetrieveStream API

The solution combines Amazon Bedrock Knowledge Bases with API Gateway, Lambda, Cognito, SQS, DynamoDB, and S3 to deliver secure, scalable document chat without building retrieval infrastructure.