Amazon Bedrock Knowledge Base adds single-shot RAG and passage-only modes

AWS said on 9 November 2026 that Amazon Bedrock Managed Knowledge Base now supports single-shot RAG (retrieval-augmented generation) and retrieval-only requests through its AgenticRetrieveStream API, alongside the existing agentic retrieval. Five AWS authors describe the change in a vendor-published post. The choice is made with one parameter, maxAgentIteration, whose range is now 0 to 5.

Three modes, one parameter

• maxAgentIteration=0: one retrieval, no generation and zero foundation-model calls. The post says it matches calling the Retrieve API directly, but passages arrive without score or location fields.
• maxAgentIteration=1: one retrieval, then one answer-generation call, with no query rewriting or planning. AWS suggests this as the default if most traffic is well-scoped, direct questions.
• maxAgentIteration=2 to 5: agentic mode, unchanged. A model splits the question into sub-queries, checks whether the evidence is sufficient and searches again, for up to n minus 1 planning rounds.

One iteration equals one foundation-model call, and the last one is always the answer. The initial speculative retrieval does not count against the limit. All three modes work across multiple knowledge bases.

What AWS shows

The example is a fictional bank assistant. A question about which documents a sole proprietor needs uses value 1 because one checklist holds the answer. A comparison of requirements for sole proprietors and LLCs, plus checks for nonresident owners, uses 4. An employee who wants the policy text itself gets value 0 with generation off. The post stresses that choosing the mode is application logic, not automatic routing.

What the post leaves out

It gives no latency, cost or accuracy numbers and says its examples are not a customer deployment or measured results. It says single-shot RAG can reduce latency and model usage compared with agentic retrieval but does not say by how much. It also does not say how an application should decide which questions need planning.

What a team can take from it

AWS advises building a set of representative questions and expected answers from your own documents, running single-shot and your current agentic setting on the same knowledge base, and measuring latency, request cost and whether answers are supported by the retrieved passages. Include questions that span several documents and questions the documents cannot answer. Then pick the lowest iteration limit that meets each workflow's quality bar.