AWS ships CDK app for agentic retrieval, one 'cdk deploy --all'

The starting point for this post is a familiar limitation: classic Retrieval Augmented Generation does one retrieval and one generation, so questions whose answer spans two document sets, or that need a second look after the first result, tend to come back fluent but incomplete. Putting a reasoning agent in the loop changes that, because the agent decides whether to retrieve, which knowledge base to consult, and whether to go back for more before composing an answer.

Amazon Bedrock's move to a Managed Knowledge Base is what makes that pattern deployable rather than bespoke. Ingestion, storage, indexing, and retrieval are handled by the service, and agentic retrieval plus the AgentCore Gateway connector are offered only on the managed variant — so the choice of foundation is effectively made for you if you want this capability. AWS notes the customer-managed option still fits teams that want to operate their own vector store, for example to meet data-residency or compliance requirements.

The design choice worth noticing is splitting the two corpora into two separate knowledge bases, each exposed as its own tool. That makes routing an explicit, observable decision: the span tree shows which tool the agent picked, and retrieval-quality metrics stay separate per knowledge base instead of averaging across unrelated domains. Roughly half of what the application provisions exists for that reason — to report what the agent did, not just to answer the question.