AWS Bedrock Managed Knowledge Base: single-shot and agentic retrieval modes
AWS says Amazon Bedrock Managed Knowledge Base now lets teams choose between single-shot RAG, retrieval-only output, and fuller agentic retrieval by setting maxAgentIteration and generateResponse per request.
What happened
The AWS Machine Learning Blog describes three retrieval workflows for Amazon Bedrock Managed Knowledge Base controlled by maxAgentIteration. It says maxAgentIteration=1 is single-shot RAG: one retrieval, then one grounded answer. It says maxAgentIteration=0 is retrieval-only and, with generateResponse=false, returns passages without language model calls, equivalent to the Retrieve API. AWS also says agentic mode adds query decomposition, evidence evaluation, and iterative searches, and that the choice of mode is application logic rather than automatic routing. The post further advises teams to compare latency, cost, and answer quality on representative questions before selecting the lowest iteration limit that still meets requirements.
Why it matters
AWS’s split between single-shot, retrieval-only, and fuller agentic retrieval gives product teams a per-request control for trading latency, cost, and answer completeness. The practical decision is whether a workflow can tolerate one retrieval or needs iterative search, but the evidence is AWS’s own API description, not independent benchmarking.
Engineering impact
Engineering teams can keep one retrieval stack and vary behavior through maxAgentIteration and generateResponse instead of maintaining separate endpoints for passages and generated answers. The key constraint is that mode choice remains application logic, so engineers must wire policy decisions, tests, and fallbacks around the parameter values rather than expect automatic routing.
Leadership impact
For platform and product leaders, the main organizational implication is a governance choice: approve the lowest retrieval effort that still meets quality requirements for each workflow. AWS explicitly recommends comparing latency, cost, and supported answers on representative questions, including hard and unanswerable cases, before standardizing the default.
Editorial judgment
This changes a practical platform decision: teams can tune retrieval effort per request instead of treating agentic search as the default. The source is AWS’s own description, so the reliable takeaway is a scoped implementation choice, not independent proof that the modes are generally better.
Watch next
The next decision is whether a team’s questions are mostly direct enough for maxAgentIteration=1 or require iterative search. A useful evaluation benchmark is whether answers stay supported when questions span multiple documents. The source does not show independent performance results, so broader rollout criteria remain unknown outside AWS’s description.
Limitations & uncertainty
All implications come from AWS’s own blog post about its Managed Knowledge Base API. The source describes intended behavior and suggested evaluation criteria, but it does not provide independent third-party validation, benchmark data across workloads, or evidence that these modes generalize beyond the documented service context.