Amazon Web Services (AWS) has introduced a new approach to enterprise agentic retrieval using a Managed Amazon Bedrock Knowledge Base, leveraging the AWS Cloud Development Kit (AWS CDK). Traditional Retrieval Augmented Generation (RAG) systems often fail when questions require information from multiple sources or need iterative retrieval. The new solution addresses these limitations by integrating a reasoning agent that decides retrieval steps, consults relevant knowledge bases, and composes grounded, cited answers.
The Managed Knowledge Base simplifies operations by handling ingestion, storage, indexing, and retrieval automatically. This eliminates the need for separate OpenSearch Serverless collections and reduces operational complexity. The AgenticRetrieveStream API now offers multi-turn planning, tool-driven retrieval, and synthesized answers as a first-class service feature.
Observability and evaluation are built into the design, with Amazon CloudWatch dashboards providing telemetry across seven layers. This ensures transparency in the agent's decision-making process, which is crucial for production environments. The solution deploys as a single AWS CDK application in Python, streamlining the setup process.
For those familiar with the 2024 RAG solution, the new system replaces the vector store with agentic retrieval and adds extensive monitoring capabilities. The Managed Knowledge Base is the only option supporting agentic retrieval, making it the foundation for this advanced capability.