LangSmith Adds OpenTelemetry Support for Unified LLM and System Telemetry
Set OTEL_EXPORTER_OTLP_ENDPOINT to https://api.smith.langchain.com/otel and configure headers to ingest traces into LangSmith.
Configure your OpenTelemetry exporter to point to https://api.smith.langchain.com/otel with the appropriate API key and project headers to ingest traces into LangSmith.
Summary
LangSmith’s new release introduces native ingestion of OpenTelemetry traces, allowing developers to export telemetry data from any language or framework that supports the OTEL spec. The platform now accepts traces following the OpenLLMetry semantic convention, which is designed for generative AI workloads and covers LLM models, vector databases, and common frameworks. A dedicated OTEL endpoint (https://api.smith.langchain.com/otel) is provided, and developers can configure their exporters with headers that include the LangSmith API key and project name. The update is demonstrated with three example setups: a Python client using the OpenTelemetry SDK, the Traceloop SDK’s OpenLLMetry integration, and the Vercel AI SDK’s client‑side exporter. Each example shows how to instrument an OpenAI chat completion and view the resulting trace in the LangSmith dashboard. Future plans include support for the evolving OpenTelemetry Gen AI semantic convention, expanding the range of telemetry that can be ingested.
This feature enables teams to unify LLM monitoring with system telemetry, providing a single view of application performance and agent behavior.
The release is a significant step toward comprehensive observability for AI‑powered applications.
Key changes
- LangSmith now ingests OpenTelemetry traces via a dedicated OTEL endpoint
- Supports OpenLLMetry semantic convention for generative AI workloads
- Provides example setups for Python OpenTelemetry SDK, Traceloop SDK, and Vercel AI SDK
- Future support planned for OpenTelemetry Gen AI semantic convention
- Allows tracing of LLM calls, vector DB interactions, and common AI frameworks