Amazon Bedrock Managed Knowledge Base Adds Single-Shot RAG and Retrieval-Only Modes, Letting Developers Match Retrieval to Question Difficulty

What happened

In an official blog post, AWS announced that Amazon Bedrock Managed Knowledge Base now supports single-shot retrieval-augmented generation (RAG) and retrieval-only requests through AgenticRetrieveStream, an API (the interface applications use to send requests to the service). RAG means first finding relevant passages in a company's documents, then having an AI model write an answer based on those passages. A knowledge base is the store that holds those documents for the AI to query.

According to AWS, single-shot RAG searches the knowledge base once, and a foundation model (the large AI model that understands the question and generates text) then writes an answer from the passages found. Retrieval-only returns just the matching passages without generating an answer, so an application or user can work with the original text directly. These two modes join the existing agentic retrieval, in which the service plans and runs multiple searches on its own before answering.

AWS explains that in agentic retrieval, the foundation model breaks the question into sub-queries, searches the relevant knowledge bases and evaluates whether the results are sufficient. When necessary, it adjusts the search within the configured iteration limit, and it can also fetch the full document when a passage lacks context. When response generation is enabled, the service combines the retrieved information into an answer with citations, and trace events (a step-by-step log for developers) show the planning and retrieval steps.

How to choose among the three modes

AWS says the range of maxAgentIteration, a setting in each request, is now 0 to 5, and this single parameter decides which mode each request uses. One iteration equals one foundation model call. A value of n allows at most n minus 1 rounds of agentic planning, with the final iteration reserved for generating the answer. AWS also notes that the initial speculative retrieval run before the first planning step does not count toward the iteration total.

Practical impact for users and developers

For internal company Q&A assistants, simple lookups and complex comparisons can now share the same API. AWS notes that single-shot RAG uses just one foundation model call, which can reduce latency and model usage compared with agentic retrieval, while retrieval-only makes no foundation model call at all.

AWS gives three usage examples. For the question "What documents are required to open a business checking account for a sole proprietor?", use maxAgentIteration=1 with response generation on. For a question comparing account-opening requirements for sole proprietorships and limited liability companies (LLCs), use maxAgentIteration=4. When someone needs to read the original policy text, use 0 with response generation off.

Limits and caveats

• AWS says choosing a mode is application logic; the parameter does not route requests automatically.
• With maxAgentIteration=0, setting generateResponse to true is an invalid configuration and returns a 4xx error (the service rejects the request).
• Retrieval-only passages appear in the stream's result events and do not include score or location fields, unlike the response format of a direct Retrieve call.
• AWS says its example assistant does not approve accounts and does not replace the bank's review and authorization processes.
• AWS cautions that a higher iteration limit does not always produce better answers, and recommends evaluating with representative questions.

How to get started (per AWS)

1. Prepare a set of representative questions and expected answers from your own documents, including questions that require multiple documents to answer and questions the documents cannot answer.
2. Following the AgenticRetrieveStream API reference, set maxAgentIteration and generateResponse in the request.
3. On the same knowledge base, compare single-shot RAG with your existing agentic setup, measuring response latency and request cost, and checking whether answers are correct and supported by retrieved passages.
4. Based on the results, choose the lowest iteration limit that meets the quality requirements for each workflow.

Browse the latest news in this topic

• AI news