ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Amazon Web Services: Switching Knowledge Retrieval Depth Between Pure Retrieval, Single-Shot Generation, and Multi-Turn Planning with One Parameter

The Amazon Web Services Machine Learning blog explains how the AgenticRetrieveStream interface of Amazon Bedrock Managed Knowledge Base switches between three retrieval modes via the maxAgentIteration parameter: 0 for pure retrieval with zero LLM calls, 1 for single-shot retrieval plus one generation, and 2 to 5 for multi-turn planning. The article presents the core model for counting iterations—one iteration equals one foundation model call, with the loop always ending in a response generation call—and explains how this model makes behavior predictable.

2026-11-10 04:13 GMT+8

When building knowledge base question-answering applications, developers face a common contradiction: simple lookups require speed and low cost, while complex comparisons require multi-turn searches and evidence integration. If separate systems are built for each type of request, maintenance costs double; if multi-turn planning is used uniformly, simple queries incur unnecessary latency and model invocation fees. An article published by the Amazon Web Services Machine Learning blog on November 9, 2026, explains how Amazon Bedrock Managed Knowledge Base resolves this issue through the maxAgentIteration parameter of the AgenticRetrieveStream interface: the same API, one integer, three behaviors.

Understanding this parameter requires clarifying how iterations are counted. The model provided in the article is: one iteration equals one foundation model call, and the loop always ends with a call to generate an answer. Therefore, a value of n means up to n minus 1 rounds of planning, with the last round reserved for generating the response. Each planning round consists of one query rewrite plus one retrieval. There is also an exception that does not consume iteration quota: speculative retrieval—the initial retrieval launched before the first planning step does not count against maxAgentIteration; it only serves to make requests faster. This counting model is the most operationally valuable part of the article because it allows developers to predict exactly how many model invocations a request will trigger at most.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Each of the three tiers has its applicable scenarios. When maxAgentIteration is set to 0, one retrieval is executed, no answer is generated, and no foundation models are invoked, which is equivalent to directly calling the Retrieve API. This is suitable for callers who have already standardized on AgenticRetrieveStream but only need to return original text passages. When set to 1, one retrieval plus one answer generation is executed, representing classic single-shot RAG (Retrieval-Augmented Generation), with no query rewriting or planning rounds. This is suitable for direct questions where answers are concentrated in a single document or checklist. When set to 2 to 5, the system enters agentic mode: the foundation model decomposes complex questions into sub-queries, retrieves across multiple knowledge bases, evaluates whether the evidence is sufficient, and refines the search to continue retrieving if necessary. The article suggests that if most traffic consists of well-scoped direct questions, 1 is an appropriate default value.

It should be noted that the article itself is a vendor's explanation of its own API behavior and provides no measured latency, cost comparisons, or independent benchmark tests. The banking example in the article is explicitly marked as fictional illustration and does not represent customer deployments or measurement results. Mode selection is described in the article as application logic rather than automatic API routing—developers must still determine which tier to use for each request. The adoption path suggested by the article is: compare single-shot RAG with your current agentic configuration using a representative set of questions, measure latency and cost, verify that answers are correct and supported by retrieved passages, and then select the lowest iteration limit that meets quality requirements for each workflow.

Sources

  • aws.amazon.comAmazon Web Services↗
Continue researching →