On October 26, 2026, AWS published a technical blog post officially launching an agentic retrieval reference implementation based on the AWS Cloud Development Kit (CDK), specifically designed for the Amazon Bedrock Managed Knowledge Base (Type: MANAGED). This solution was introduced in August 2024 alongside the Managed KB and, unlike traditional single‑call RAG approaches, supports multi‑turn, tool‑driven, semantic routing across multiple knowledge bases.
This CDK application uniformly deploys four tightly coupled CloudFormation stacks: -kb (including S3 and two managed knowledge bases), -gateway (configured with AgenticRetrieveStream targets and MCP authentication), -agent (an Instrumented Strands agent based on the AgentCore Runtime, built via CodeBuild), and -dashboards (two sets of CloudWatch dashboards covering seven layers of observability). All inter-stack dependencies—totaling seven cross-stack references, such as KB IDs and Gateway ARNs—are automatically resolved by the CDK to determine the deployment order, eliminating the need for manual export or lookup.
The solution employs two types of prebuilt corpora: synthetic financial data (a fictional Octank company 10‑K report, CC0 licensed) and real U.S. government tornado reports (IF12695, public domain). Both are automatically parsed, chunked, and embedded—using Bedrock’s hosted models—and indexed by a managed knowledge base, eliminating the need for users to manage vector storage or pay for idle OCU costs. Semantic routing explicitly selects two independent AgenticRetrieveStream tools (corresponding to the financial and weather knowledge bases, respectively), ensuring observability of domain‑isolated metrics.
Observability is built into the architecture: Layer 7 telemetry (L1–L7) covers native metrics from the knowledge base, ingestion, retrieval quality (without citation-based computation), Gateway/MCP calls, OpenTelemetry span trees, token usage, and evaluation scores; with L1/L4/L5 automatically reported, while L3/L6/L7 are published as custom CloudWatch metrics. Evaluation supports two modes—on-demand (notebook-driven LLM-as-judge) and continuous (online sampling sessions, including four built-in evaluators such as Correctness)—with the demo defaulting to 100% sampling; in production, the sampling rate should be adjusted based on cost considerations.
Deployment requires meeting several prerequisites: the target region (default: us-west-2) must support Managed KB, AgenticRetrieveStream, and AgentCore Runtime/Gateway; you must have permissions for services such as Bedrock, AgentCore, IAM, CloudWatch, and X-Ray; CloudWatch Transaction Search must be enabled; CDK must be initialized; and the runtime environment must be Node.js 18+ or Python 3.13+, with no local Docker required. Cleanup is performed using `cdk destroy --all`, and S3 buckets and ECR repositories are configured with a `removal_policy` to enable automatic deletion. All code and test notebooks are open-source on GitHub.