AWS outlined how LangChain applications can use Bedrock’s agentic retrieval to improve evidence coverage for complex questions, with added cost and latency.
AWS published a technical guide on October 5, 2026, showing how developers can compare standard and agentic retrieval in Amazon Bedrock Managed Knowledge Bases using LangChain. The walkthrough focuses on questions with multiple parts, where a single similarity search may return relevant passages but leave some requested details unsupported.
The standard Retrieve API performs one hybrid search and returns scored document chunks. AgenticRetrieveStream instead runs a planning loop: it can split a query into sub-questions, retrieve evidence for them, assess whether that evidence is sufficient and search again. AWS exposes agentic retrieval through the langchain-aws package as a standalone function, rather than as a switch on the standard LangChain retriever. Developers who need to inspect the plan can call the streaming API directly and examine its trace events.
AWS’s example asks for a comparison of two services across three dimensions, amounting to six sub-intents. With five retrieved chunks, the standard path covered four of the six; ten chunks covered all six, but included duplicated and irrelevant material. In AWS’s MuSiQue benchmark evaluation, agentic retrieval improved recall over single-shot retrieval, with the largest gains on harder multi-hop questions. Gains for single-hop questions were under five points.
The choice has operational consequences. AWS describes standard retrieval as cheaper and faster, making it a better fit for short, focused questions. Agentic retrieval makes multiple model calls and typically adds cost and latency, but can help with comparative, exploratory or multi-part queries, including questions spanning multiple knowledge bases. The guide recommends routing based on query complexity rather than sending every request through the planning loop.
The walkthrough also flags integration details: agentic retrieval is a function rather than a conventional LangChain retriever, and the trace events require a direct Boto3 call. The sample requires Python 3.12 or later, langchain-aws 1.6.3 or later, LangChain 1.0 or later, and Boto3 1.43.32 or later. AWS says teams should evaluate their own query mix before deciding how often to use the higher-cost path.