Agentic Retrieval for Amazon Bedrock Managed Knowledge Base

AWS has announced a new agentic retrieval mode for Amazon Bedrock Managed Knowledge Bases, available via the AgenticRetrieveStream API. Unlike the classic single-shot Retrieve, which struggles with multi-part, comparative, and research questions, agentic retrieval uses a planning loop based on a foundation model: the model breaks down the query into sub-queries, searches each one, evaluates whether enough facts have been found, and repeats iterations if necessary. The result is a stream of trace events, including planning steps, search steps, and final deduplicated chunks, along with a generated reasoned response when appropriate. The API supports up to five knowledge bases, with query routing based on their textual descriptions. Configuration parameters include maxAgentIteration (maximum iterations) and the choice of planner model. Pricing is based on planner calls plus internal Retrieve calls. On the MuSiQue benchmark, agentic retrieval showed an absolute recall improvement of 20% compared to single-shot retrieval, with the greatest effect on the most complex questions (up to +37.3% for 4-step questions). Practical recommendations include using agentic retrieval for complex queries and Retrieve for simple ones, along with mandatory logging of trace events in Amazon CloudWatch.