Instrumenting Amazon Bedrock, Bedrock Agents, and AgentCore with OpenTelemetry

Amazon Bedrock can be instrumented with OpenTelemetry by applying gen_ai.* semantic conventions across three different access points: raw model calls via boto3, managed Bedrock Agents through the InvokeAgent API, and custom agents on Bedrock AgentCore. These conventions ensure uniform telemetry data collection, allowing for a consolidated dashboard that tracks spend and latency across all Bedrock interactions by directing outputs to an OTLP backend like OpenObserve. The unified telemetry approach replaces the deprecated gen_ai.system attribute with gen_ai.provider.name, focusing instead on model IDs for consistent data analysis. This setup supports cost tracking by converting token usage into dollar amounts using a custom price map, while maintaining privacy by excluding prompt content from telemetry unless explicitly opted in. OpenTelemetry's integration with OpenObserve facilitates comprehensive trace, metric, and log management, enabling users to monitor operational efficiencies and costs associated with Amazon Bedrock deployments effectively.