langsmith_tracing

LangSmith Tracing

This sample demonstrates LangSmith tracing integration with Temporal workflows using the LangSmithPlugin.

Two examples are included:

• basic/ — A one-shot LLM workflow that sends a prompt to OpenAI and returns the response.
• chatbot/ — A long-running conversational workflow with tool calls (save/read notes) and update handlers.

Prerequisites

Install dependencies:

uv sync --group langsmith-tracing

Set environment variables:

export OPENAI_API_KEY="sk-..."
export LANGSMITH_API_KEY="lsv2_..."
export LANGCHAIN_TRACING_V2=true

A local Temporal server must be running (temporal server start-dev).

Three Layers of Tracing

This sample shows three complementary ways LangSmith captures trace data:

1. Automatic (wrap_openai) — Wrapping the OpenAI client with wrap_openai() automatically creates a child span for every LLM call, capturing model parameters, token usage, and latency. No extra code needed.

2. Explicit (@traceable) — Decorating functions with @traceable creates named spans for your business logic. You control the name, tags, metadata, and run_type (chain, llm, tool, retriever). This is how you structure traces to tell a story about what your application is doing.

3. Temporal (add_temporal_runs=True) — The LangSmithPlugin can optionally create LangSmith runs for each Temporal workflow execution and activity execution, giving visibility into the orchestration layer alongside your LLM calls.

add_temporal_runs

By default, LangSmithPlugin(add_temporal_runs=False) only propagates LangSmith context so that @traceable and wrap_openai calls nest correctly.

Set add_temporal_runs=True to also create LangSmith runs for Temporal operations (workflow executions, activity executions, signals, etc.), giving full visibility into the orchestration layer. Both examples support a --add-temporal-runs CLI flag to toggle this.

Further Reading

• LangSmith documentation
• Temporal Python SDK LangSmith plugin
• LangSmith @traceable guide
• LangSmith wrap_openai guide