On November 9, 2026, AWS announced that Amazon Bedrock Managed Knowledge Base would support retrieval-only and single-shot retrieval-augmented generation (RAG) through the AgenticRetrieveStream API, alongside multi-step agentic retrieval. The application selects a mode for each request with maxAgentIteration; the setting does not route questions automatically. AWS’s announcement
Choose retrieval depth per request: return passages for your application to handle, use one retrieval for a direct question, or allow multiple retrieval steps when the answer may depend on several searches.
What each maxAgentIteration value does
The setting ranges from 0 to 5. At 0, Bedrock retrieves passages without calling a foundation model or generating an answer, so generateResponse must be false. At 1, it retrieves once and generates an answer without planning rounds. Values 2–5 retain multi-step agentic retrieval.
| Mode | maxAgentIteration | What happens | A suitable fit |
| Retrieval-only | 0 | One retrieval returns passages; no foundation-model call or generated answer. | Your application will display or process the passages itself. |
| Single-shot RAG | 1 | One retrieval is followed by a generated answer, with no planning rounds. | A direct question likely to be answered by one retrieval. |
| Agentic retrieval | 2–5 | The model can plan searches, retrieve for subqueries, and assess whether the evidence is sufficient before answering. | A complex question that may need evidence from multiple documents. |
If generateResponse is enabled at 0, the request returns a 4xx error. For values 2–5 with generation enabled, the final model call is reserved for the answer, leaving up to n−1 planning rounds. AWS counts one model call as one iteration; an initial speculative retrieval happens before planning and does not count toward the limit. A single-shot request can query multiple knowledge bases: “single-shot” means one retrieval, not one knowledge base. AWS documents the mode behavior and iteration limit.
What the API returns and what integration requires
AgenticRetrieveStream requires an agenticRetrieveConfiguration, at least one message, and at least one retriever. Its stream can return retrieval results, trace events, and a generated response when generation is enabled. Retrieval-only passages arrive in the result event; AWS says this output does not include the score or location fields. A generated answer can include citations that map answer spans to retrieved results. The API reference describes the request and response structure.
AWS’s guide describes agentic retrieval with fully managed Bedrock knowledge bases and up to five retrievers per request. Its permissions example lists bedrock:AgenticRetrieveStream, bedrock:Retrieve, bedrock:GetDocumentContent, and bedrock:InvokeModelWithResponseStream. That example concerns agentic retrieval; it does not change the retrieval-only mode’s behavior at maxAgentIteration=0, which makes no foundation-model call. AWS’s user guide covers the retrieval requirements and permissions.
How to evaluate a mode for your workload
AWS recommends testing representative questions and expected answers against the same knowledge base. Include questions that require multiple documents and questions the knowledge base cannot answer. Compare latency and request cost, then check whether answers are correct and supported by the passages retrieved. That makes the tradeoff concrete: a lower iteration limit can suit simpler questions, while questions needing repeated searches may benefit from agentic retrieval.
AWS says single-shot RAG can reduce latency and model usage compared with agentic retrieval, but describes that as a potential benefit rather than reporting a quantified result. The right setting depends on how your own representative questions perform.
For a wrong answer despite having the right document, commenters in an RAG troubleshooting discussion suggest giving the model the relevant passage directly with the same prompt, then comparing that response with the context sent by the application. Missing headings or table context can also be worth checking; this is general community troubleshooting advice, not a finding about Bedrock’s new modes.