Amazon's New AI Search Answers Your Multi-Part Questions Better

When you ask an AI assistant a multi-part question—like "compare two products across three features"—standard search often fails. It uses one search query to cover all your intents, but ends up with chunks that only partially answer you. Amazon's new agentic retrieval fixes this by planning: it splits your question into sub-queries, runs them, checks if it has enough evidence, and searches again if needed. This means you get a more complete answer without having to rephrase or ask again.

Amazon Bedrock Managed Knowledge Base handles all the technical stuff—chunking, embedding, storing, and retrieving documents. You just point it to your data (like an S3 bucket), and it does the rest. The new AgenticRetrieveStream API streams back the steps it took, so you can see how it arrived at the answer. The older Retrieve API does one hybrid search and returns scored chunks. Both are available in the langchain-aws package.

But agentic retrieval isn't always the best choice. It costs more because it runs multiple searches and calls extra APIs like GetDocumentContent. For simple questions, the standard single-shot retrieval is cheaper and faster. The key is knowing when to use which: agentic for complex, multi-part questions; standard for straightforward lookups. This matters because it affects your cloud bill and response time.

For everyday users, this means AI assistants built on Amazon Bedrock can now give you better answers to complicated questions—like planning a trip with multiple constraints or comparing insurance plans. It reduces the back-and-forth and makes AI more reliable for real-world tasks. Just remember: smarter search costs more, so businesses will need to balance quality and cost.