AWS has published guidance for agentic retrieval in Amazon Bedrock Managed Knowledge Bases, aimed at questions that are too complex for a single similarity search. The new AgenticRetrieveStream API can plan and iterate over retrieval steps and can generate a response in the same call.
The problem is familiar to teams building retrieval-augmented generation systems. A query such as “Compare our 2020 and 2023 strategy” or “What are the three biggest risks across product lines?” does not map cleanly to one embedding. Standard top-k retrieval can average together competing intents and return chunks that are related but not useful.
Agentic retrieval breaks the question into steps, inspects intermediate results, and uses traces so developers can see how the system worked through the task. AWS positions it as a better fit for multi-part, comparative, and exploratory questions spanning documents, tickets, transcripts, and web content.
The tradeoff is complexity. Simple lookups may still be better served by the standard Retrieve API, while agentic retrieval is most useful when the question requires planning and evidence gathering across several pieces of information.