Amazon Bedrock launches agentic retrieval for complex multi-part queries

Classic single-shot retrieval fails on multi-part questions like "Compare our 2020 and 2023 strategy." A single embedding can't represent multiple intents, so top-k results mix competing sub-intents. As shown with shareholder letters, the retriever returns scattered or dominated chunks, missing context entirely. Analysts must manually decompose queries, running separate searches—a time-consuming process.

Amazon's solution is agentic retrieval for Bedrock Managed Knowledge Bases, available via the AgenticRetrieveStream API. It runs a foundation model-driven planning loop: the model decomposes the question, retrieves against each part, evaluates if evidence suffices, and iterates as needed. It can generate a grounded response in the same call or return pure retrieval results (generateResponse=False). Every step (SpeculativeRetrieval, Planning, Execution) streams trace events, giving full transparency into the reasoning. This automates what a human analyst would do, making complex research tasks—like comparing strategy across years or risks across product lines—instant and accurate.