Build agentic retrieval on a Managed Amazon Bedrock Knowledge Base with CDK

This article demonstrates building enterprise agentic retrieval on Amazon Bedrock Managed Knowledge Bases with AWS CDK, featuring multi-KB semantic routing, a reasoning agent, and seven layers of observability with on-demand and continuous evaluation.

• Managed Knowledge Bases eliminate vector store provisioning and scaling, with Amazon Bedrock handling ingestion, storage, indexing, and retrieval automatically
• AgenticRetrieveStream enables multi-turn planning, semantic routing across knowledge bases, and synthesized cited answers instead of single-shot retrieval
• AgentCore Gateway exposes each knowledge base as an MCP tool, allowing agents to route questions to the appropriate corpus and retrieve iteratively
• Seven observability layers span KB metrics, ingestion, retrieval quality, Gateway/MCP metrics, OTEL spans, token usage, and evaluation scores
• On-demand evaluation uses LLM-as-judge over session spans; continuous evaluation samples live sessions automatically with built-in evaluators
• Single CDK application deploys four stacks (kb, gateway, agent, dashboards) with automatic dependency ordering and cross-stack value threading
• Two CloudWatch dashboards provide end-to-end agentic observability and per-KB operational metrics including index size, retrieve volume, and token usage

The solution demonstrates how managed infrastructure and agentic capabilities eliminate operational overhead while providing comprehensive observability for production enterprise retrieval systems.