Amazon Bedrock Managed Knowledge Base Adds Agentic Retrieval with LangChain Integration

What changed?

Amazon Bedrock Managed Knowledge Base now provides agentic retrieval capabilities. Developers can access two retrieval APIs: Retrieve (standard, single-shot search) and AgenticRetrieveStream (multi-step, agentic retrieval) through the langchain-aws package. Agentic retrieval enables the system to break down complex queries into sub-queries, execute each, judge evidence sufficiency, and loop back to re-search if needed. This feature is available in supported AWS regions such as US East (N. Virginia).

Why does it matter to an everyday developer?

Standard retrieval methods use a single vector search and often provide only partial answers to multi-intent or complex questions. This can limit the effectiveness of RAG applications when handling queries comprising several intent clusters (e.g., comparing products across multiple dimensions). With agentic retrieval, developers can now offer more complete and accurate results for these multi-part queries without manual orchestration. The managed service eliminates the need to set up or operate a vector store, embeddings, or re-ranking models, simplifying MLOps for RAG applications.

What can the developer do now?

Developers building RAG applications on Amazon Bedrock can choose between standard and agentic retrieval modes through the langchain-aws package. They can stream trace events to analyze how queries are planned and executed, compare effectiveness and cost of retrieval paths, and build applications that more reliably answer complex, multi-faceted questions. Setup requires using Python 3.12+, appropriate versioned dependencies, access to a supported AWS region, and permissions as outlined in the official walkthrough. Existing Bedrock customers can experiment with agentic retrieval without migrating their storage infrastructure.