AWS adds single-shot RAG and retrieval-only modes to Bedrock
Amazon Bedrock Managed Knowledge Base now supports single-shot Retrieval Augmented Generation and retrieval-only requests through AgenticRetrieveStream. AWS says the expanded controls let applications match retrieval effort to the complexity of each question.
- Amazon Bedrock Managed Knowledge Base now supports single-shot RAG and retrieval-only requests.
- maxAgentIteration=0 returns passages without making foundation model calls.
- maxAgentIteration=1 performs one retrieval followed by answer generation.
- Values from 2 to 5 provide agentic planning and iterative retrieval.
- AWS illustrates the modes with an internal banking policy assistant.
Three retrieval modes through one API
The service now lets developers choose among three modes by setting maxAgentIteration. Retrieval-only requests return matching passages without generating an answer, allowing an application or user to work directly with the source text. Single-shot RAG searches a knowledge base once and uses a foundation model to generate an answer from the retrieved passages.
Agentic retrieval remains available for more complex requests. In this mode, a foundation model breaks a question into sub-queries, searches relevant knowledge bases, evaluates the evidence and can refine its searches when information is missing. When response generation is enabled, the resulting answer can include citations, while trace events show the planning and retrieval steps.
AWS presents the same API as a way to handle both direct lookups and questions that require comparisons across documents, without building separate systems for each type of request.
What the maxAgentIteration values do
The available range is now 0, 5, and all three modes support multiple knowledge bases. A value of 0 performs one guaranteed retrieval and generates no answer. AWS says no foundation model is invoked in this mode, which makes it equivalent to calling the Retrieve API directly. Setting generateResponse=true with this value is invalid and returns a 4xx error.
A value of 1 enables single-shot RAG: the service retrieves once and then generates an answer, with no agentic planning or query rewriting. AWS describes this as the setting for direct, well-scoped questions and says it can reduce latency and model usage compared with agentic retrieval.
Values from 2 to 5 retain agentic behavior. With response generation enabled, a value of n allows up to n minus 1 planning rounds before the final answer-generation call. Each planning round consists of a query rewrite and a retrieval. AWS says lower values provide faster and more predictable calls, while higher values give the planner more room to decompose and refine difficult questions.
Designed for direct lookups and complex comparisons
AWS illustrates the feature with an internal banking policy assistant. An employee asking which documents are needed to open a business account could use single-shot RAG for a checklist lookup. A question comparing requirements for sole proprietorships, LLCs and partnerships with multiple owners could use agentic retrieval, because the information is spread across several policy guides.
The company says agentic retrieval can support multi-part questions by running targeted searches, perform follow-up searches when evidence is missing and retrieve a full document when a passage lacks enough context. Retrieval-only mode is intended for cases in which an employee needs to review the source passages instead of receiving a generated answer.
The examples are illustrative and do not describe a customer deployment or measured results. For application builders, the change provides a parameter-based way to trade speed and model usage against the additional planning that complex questions may require, while continuing to use AgenticRetrieveStream.
