The Pulse

AWS adds single-shot and passage-only retrieval to Bedrock

Amazon Bedrock Managed Knowledge Base now supports single-shot answer generation and retrieval without a generated answer through AgenticRetrieveStream. AWS says developers can choose among those modes and agentic retrieval using one parame

AI.info Team · Written with AI · fact-checked · 2026-11-09

One API, three retrieval modes

AWS says Amazon Bedrock Managed Knowledge Base now supports single-shot Retrieval Augmented Generation (RAG) and retrieval-only requests through AgenticRetrieveStream. The new options let an application retrieve passages and either generate an answer from them or return the source text without generating one. The service also retains agentic retrieval, which plans and runs multiple searches before producing an answer. AWS describes the change as a way to select how much retrieval work to apply to each request through the same API.

The key setting is maxAgentIteration, whose supported range is now 0–5. AWS says a value of 0 returns passages without making a foundation-model call, while 1 performs one retrieval and generates an answer. Values from 2 to 5 retain agentic retrieval, allowing up to n minus 1 planning rounds before answer generation. The company says each planning round includes a query rewrite and a retrieval, with the final iteration reserved for generating the response. AWS says a speculative retrieval that runs ahead of the first planning step does not count toward maxAgentIteration.

Choosing how much to retrieve

At maxAgentIteration=0, the request must set generateResponse to false; AWS says setting it to true returns a 4xx error. The company says maxAgentIteration=0 is equivalent to calling the Retrieve API directly, though the response format differs. With maxAgentIteration=1 and response generation enabled, AWS says the service retrieves once and generates a grounded answer without planning or query rewriting. AWS says single-shot RAG uses one foundation-model call to generate the answer, which can reduce latency and model usage compared with agentic retrieval.

The post says all three modes support multiple knowledge bases, including settings of 0 and 1. For more complex questions, agentic retrieval can break a request into sub-queries, search across knowledge bases, assess evidence and refine searches if information is missing. AWS says response generation can combine retrieved information into an answer with citations, while trace events expose planning and retrieval steps.

Examples use a fictional bank assistant

For a direct question about documents required for a sole proprietor’s business checking account, the example sets maxAgentIteration=1 and generateResponse=true. For a comparison across business structures and ownership checks, the example sets maxAgentIteration=4, permitting up to three planning rounds before an answer. AWS cautions that the suitable iteration limit should be evaluated against representative questions rather than assuming that a higher setting always improves results.

For employees who need to review policy text rather than receive a generated summary, the example sets maxAgentIteration=0 and generateResponse=false to return passages without language-model calls. The application, rather than the parameter, selects which retrieval mode to use for a request. AWS recommends testing the modes against representative questions and expected answers, comparing latency, request cost, correctness and whether retrieved passages support the responses. The post says existing requests using values from 2 to 5 retain their current behavior, allowing developers to introduce the new modes selectively.