Amazon has introduced agentic retrieval for its Bedrock Managed Knowledge Bases, designed to handle complex, multi-part questions that traditional retrieval methods struggle with. Classic single-shot retrieval often fails to provide accurate context for such queries, leading to incomplete answers and wasted time for analysts. Agentic retrieval addresses these limitations by planning and iterating over retrieval steps, delivering more precise results in a single API call. This new feature is available through the AgenticRetrieveStream API and is intended to streamline the process of answering multi-intent, comparative, and exploratory questions.
Agentic retrieval works by decomposing complex queries into sub-queries, retrieving information step-by-step, and judging whether sufficient evidence has been gathered. This iterative approach allows the system to refine results based on prior findings, ensuring a more comprehensive and accurate response. For example, when analyzing changes in Amazon’s strategy over time, agentic retrieval can break down the query into specific components, such as hiring practices, long-term investment, and customer obsession, and retrieve relevant information for each. This method avoids the pitfalls of single-shot retrieval, which often returns a mix of loosely related results or a cluster dominated by the strongest signal.
The source explains that agentic retrieval is a new retrieval mode in Amazon Bedrock Managed Knowledge Bases, available through the AgenticRetrieveStream API. It runs a planning loop driven by a foundation model, which decomposes the question, retrieves against each part, and iterates as needed. The process generates a grounded response by default in the same call, though users can opt to return retrieval results without generating a response. Each step of the loop is tracked as an ordered stream of trace events, providing visibility into the planning and retrieval process.
Source: awsml
