Amazon Bedrock Knowledge Bases now offers agentic retrieval through the `langchain-aws` package, giving developers a second path alongside conventional one-shot search. The new option targets compound questions that a single query vector may represent poorly.
Standard retrieval runs one hybrid search and returns scored passages. The `AgenticRetrieveStream` API instead asks a model to break the request into subqueries, run searches, judge whether the collected evidence is sufficient and search again when necessary. It streams the plan as trace events so an application can inspect the retrieval steps.
AWS illustrates the difference with product comparisons spanning several dimensions. One-shot search may return topically relevant passages while missing part of the request; decomposition can cover each required attribute separately. It also adds model calls, latency and retrieval cost, so simple questions may still be better served by the cheaper path.
The walkthrough requires Python 3.12 or later, `langchain-aws` 1.6.3 or later and Boto3 1.43.32 or later. Permissions need special attention: full-document expansion depends on `bedrock:GetDocumentContent`, while agentic streaming and its planner model currently require broader resource permissions. Better retrieval coverage does not guarantee a correct answer; applications still need citations, evaluation and limits on sensitive document access.