AWS has published a technical walkthrough for building an enterprise agentic retrieval system with Amazon Bedrock Knowledge Bases and the AWS Cloud Development Kit (CDK). The example routes questions across separate financial and weather knowledge bases, with each corpus ingested from Amazon S3.

According to AWS, the architecture uses Managed Knowledge Bases to handle ingestion, storage, indexing, embedding, reranking and retrieval. An Amazon Bedrock AgentCore Runtime hosts the reasoning agent, while an AgentCore Gateway exposes each knowledge base as a Model Context Protocol tool. The AgenticRetrieveStream API supports multi-turn planning, retrieval from selected knowledge bases and cited answers, AWS says.

The post contrasts this approach with conventional retrieval-augmented generation, in which a question is typically embedded, matched against one index and sent to a model for answer generation. In the described design, the agent can decide whether to retrieve, select a knowledge base and retrieve again before composing its response.

The deployment also includes two Amazon CloudWatch dashboards covering seven telemetry layers, along with on-demand and continuous evaluation. AWS says the system records agent behavior through CloudWatch and AWS X-Ray, including the reasoning and retrieval workflow.

The complete example is deployed as a single Python CDK application across four stacks. AWS says CDK derives dependencies among resources such as the knowledge bases, Gateway, runtime and dashboards. Agentic retrieval and the AgentCore Gateway integration described in the post are available only with Managed Knowledge Bases, according to AWS.