Amazon Bedrock adds single-shot and retrieval-only modes to knowledge base API
AWS expanded retrieval options in Amazon Bedrock Managed Knowledge Base, letting developers control how much work the service does per request through the maxAgentIteration parameter, which now accepts values from 0 to 5. A value of 1 produces single-shot RAG: one search followed by one foundation model call that writes an answer grounded in the passages found. A value of 0 returns matching passages and makes no model call, matching what the Retrieve API does directly. Setting generateResponse to true at 0 returns an error.
Values 2 through 5 keep the existing agentic behavior, where a model splits a question into sub-queries, searches, checks whether evidence is missing and refines before answering with citations. All three modes work across multiple knowledge bases. According to AWS, the retrieve-once-then-generate pattern was not previously expressible through this API.
Why it matters
Builders can now match retrieval effort to each question by changing one parameter instead of adding a second API or a custom orchestration loop, which cuts latency and model usage on simple lookups while keeping deeper multi-step searches for complex ones.
Sources
- PrimaryAWS Machine Learning
Written by AIToolMaps from the source above and checked against it sentence by sentence.