AWS launches agentic retrieval API for Amazon Bedrock Managed Knowledge Base

AWS has published details on Agentic Retrieval, a new capability for Amazon Bedrock Managed Knowledge Base that lets a foundation model decompose complex, multi-part questions into sub-queries, retrieve iteratively across one or more knowledge bases, and evaluate whether results are sufficient before returning an answer. The feature ships as a new API, AgenticRetrieveStream, and was detailed in a July 23, 2026 post on the AWS Machine Learning Blog by AWS engineers Omar Elkharbotly, Sailik Sengupta, and Shreya Pawaskar.

What's new

The post's core argument is that classic single-pass retrieval breaks down once a question has more than one part. As the authors put it: "This post focuses on why classic retrieval falls short on multi-part questions, how the AgenticRetrieveStream API works (including request construction and trace parsing)."

Key mechanics AWS lays out:

• Query decomposition — a foundation model breaks a compound question into individual sub-queries before retrieval starts, rather than searching on the raw question as a single string.
• Iterative retrieval — the system can run multiple retrieval passes across one or more knowledge bases, pulling in additional context when the first pass looks incomplete.
• Self-evaluation — the orchestrating model judges whether accumulated results actually answer the original question, and keeps retrieving if they don't.
• Trace output — responses include a parseable trace of the retrieval steps taken, so developers can debug why a particular answer was assembled the way it was.
• Optional synthesis — the API can return a synthesized answer using Bedrock's managed orchestration model or a model the developer specifies.

AWS frames the use cases as comparative analysis, multi-hop questions, and research-style queries — the kinds of prompts where a single retrieval pass over a vector index typically returns partial or misleading context.

Context

Amazon Bedrock Managed Knowledge Base itself is a comparatively new, fully managed retrieval-augmented-generation (RAG) service inside Bedrock, and agentic retrieval extends it rather than replacing the standard Retrieve API — AWS's guidance is explicitly about when to reach for the new agentic path versus the simpler one. The move fits a broader pattern across the major cloud AI platforms this year: retrieval is being reframed as an agentic loop (plan, retrieve, check, retrieve again) rather than a single embedding lookup, as vendors compete on how well their RAG stacks handle genuinely compound enterprise questions rather than simple lookups.

For agent builders already using Bedrock AgentCore, the post notes agentic retrieval is also exposed as a pre-built target type in the AgentCore Gateway, which cuts the integration work down to a few lines of code rather than a custom retrieval pipeline.

Why it matters

Multi-hop question answering is one of the harder unsolved problems in production RAG systems — most enterprise deployments still struggle when a user's question implicitly requires combining facts from several documents or running a comparison across sources. By pushing decomposition, iteration, and self-evaluation into a managed API rather than leaving developers to hand-roll orchestration logic on top of a vector store, AWS is competing directly on developer experience for one of the most common failure modes in applied LLM systems. It also signals where Bedrock intends to differentiate against competing knowledge-base and RAG offerings from Google Cloud and Microsoft Azure: not on raw vector search performance, but on how much of the retrieval reasoning is abstracted into the platform itself.