Amazon has enhanced its Amazon Bedrock Knowledge Base to support enterprise agentic retrieval, allowing agents to reason, retrieve across multiple knowledge bases, and generate cited answers. This solution is built using AWS CloudFormation, enabling full observability and evaluation from the start. The system automatically instruments every step with OpenTelemetry spans, providing detailed insights into the agent's operations.
The new agentic retrieval capability replaces traditional Retrieval Augmented Generation (RAG) workflows by introducing a reasoning agent that decides what to retrieve, how many times, and which knowledge base to consult. The Amazon Bedrock Managed Knowledge Base handles the ingestion, storage, indexing, and retrieval, with no need for a separate vector database. This approach supports multi-turn planning and execution, leading to more accurate and grounded answers.
The solution is designed for enterprise use, with observability and evaluation integrated into the architecture. It includes seven layers of telemetry, providing insights into latency, token usage, and answer quality. These metrics are continuously monitored and used to improve the system’s performance. The architecture uses AWS CloudFormation to deploy four native stacks, each wired to the next, ensuring a seamless and scalable deployment process.
Source: awsml
