langsmith-tracing

langsmith-tracing

Configure LangSmith observability and tracing for LLM applications built with LangChain and LangGraph frameworks.

Overview

LangSmith is the managed observability suite by LangChain that provides:

• Dashboards and alerting for LLM applications
• Human-in-the-loop evaluation capabilities
• Deep LangChain/LangGraph integration
• Run Tree model for nested traces
• MCP connectivity to Claude, VSCode

Capabilities

Core Tracing Setup

• Initialize LangSmith client and API configuration
• Configure project/workspace settings
• Set up trace collection and sampling
• Enable debug logging for agent execution

Integration Patterns

• LangChain chain tracing with automatic instrumentation
• LangGraph workflow state tracking
• Custom span creation for non-LangChain code
• Parent-child trace relationships

Debugging Features

• Fetch execution traces for analysis
• Query run history and metadata
• Export traces for offline analysis
• Compare runs across different versions

Usage

Environment Setup

# Set required environment variables
export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_API_KEY=<your-api-key>
export LANGCHAIN_PROJECT=<project-name>

Python Integration

from langsmith import Client, traceable
from langchain.callbacks.tracers import LangChainTracer

# Initialize client
client = Client()

# Use @traceable decorator for custom functions
@traceable(name="custom_operation")
def my_function(input_data):
    # Your logic here
    return result

# Initialize tracer for LangChain
tracer = LangChainTracer(project_name="my-project")

# Use with LangChain chains
chain.invoke(input, config={"callbacks": [tracer]})

Trace Retrieval

# Fetch traces from LangSmith
runs = client.list_runs(
    project_name="my-project",
    start_time=datetime.now() - timedelta(hours=24),
    execution_order=1,  # Root runs only
    error=False,  # Successful runs only
)

for run in runs:
    print(f"Run ID: {run.id}")
    print(f"Latency: {run.latency_p99}")
    print(f"Tokens: {run.total_tokens}")