langsmith-observability

LangSmith - LLM Observability Platform

Development platform for debugging, evaluating, and monitoring language models and AI applications.

When to use LangSmith

Use LangSmith when:

• Debugging LLM application issues (prompts, chains, agents)
• Evaluating model outputs systematically against datasets
• Monitoring production LLM systems
• Building regression testing for AI features
• Analyzing latency, token usage, and costs
• Collaborating on prompt engineering

Key features:

• Tracing: Capture inputs, outputs, latency for all LLM calls
• Evaluation: Systematic testing with built-in and custom evaluators
• Datasets: Create test sets from production traces or manually
• Monitoring: Track metrics, errors, and costs in production
• Integrations: Works with OpenAI, Anthropic, LangChain, LlamaIndex

Use alternatives instead:

• Weights & Biases: Deep learning experiment tracking, model training
• MLflow: General ML lifecycle, model registry focus
• Arize/WhyLabs: ML monitoring, data drift detection

Quick start

Installation

pip install langsmith

# Set environment variables
export LANGSMITH_API_KEY="your-api-key"
export LANGSMITH_TRACING=true

Basic tracing with @traceable

from langsmith import traceable
from openai import OpenAI

client = OpenAI()

@traceable
def generate_response(prompt: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}]
    )
    return response.choices[0].message.content

# Automatically traced to LangSmith
result = generate_response("What is machine learning?")

OpenAI wrapper (automatic tracing)

from langsmith.wrappers import wrap_openai
from openai import OpenAI

# Wrap client for automatic tracing
client = wrap_openai(OpenAI())

# All calls automatically traced
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}]
)

Core concepts

Runs and traces

A run is a single execution unit (LLM call, chain, tool). Runs form hierarchical traces showing the full execution flow.

from langsmith import traceable

@traceable(run_type="chain")
def process_query(query: str) -> str:
    # Parent run
    context = retrieve_context(query)  # Child run
    response = generate_answer(query, context)  # Child run
    return response

@traceable(run_type="retriever")
def retrieve_context(query: str) -> list:
    return vector_store.search(query)

@traceable(run_type="llm")
def generate_answer(query: str, context: list) -> str:
    return llm.invoke(f"Context: {context}\n\nQuestion: {query}")

Projects

Projects organize related runs. Set via environment or code:

import os
os.environ["LANGSMITH_PROJECT"] = "my-project"

# Or per-function
@traceable(project_name="my-project")
def my_function():
    pass

Client API

from langsmith import Client

client = Client()

# List runs
runs = list(client.list_runs(
    project_name="my-project",
    filter='eq(status, "success")',
    limit=100
))

# Get run details
run = client.read_run(run_id="...")

# Create feedback
client.create_feedback(
    run_id="...",
    key="correctness",
    score=0.9,
    comment="Good answer"
)

Datasets and evaluation

Create dataset

from langsmith import Client

client = Client()

# Create dataset
dataset = client.create_dataset("qa-test-set", description="QA evaluation")

# Add examples
client.create_examples(
    inputs=[
        {"question": "What is Python?"},
        {"question": "What is ML?"}
    ],
    outputs=[
        {"answer": "A programming language"},
        {"answer": "Machine learning"}
    ],
    dataset_id=dataset.id
)

Run evaluation

from langsmith import evaluate

def my_model(inputs: dict) -> dict:
    # Your model logic
    return {"answer": generate_answer(inputs["question"])}

def correctness_evaluator(run, example):
    prediction = run.outputs["answer"]
    reference = example.outputs["answer"]
    score = 1.0 if reference.lower() in prediction.lower() else 0.0
    return {"key": "correctness", "score": score}

results = evaluate(
    my_model,
    data="qa-test-set",
    evaluators=[correctness_evaluator],
    experiment_prefix="v1"
)

print(f"Average score: {results.aggregate_metrics['correctness']}")

Built-in evaluators

from langsmith.evaluation import LangChainStringEvaluator

# Use LangChain evaluators
results = evaluate(
    my_model,
    data="qa-test-set",
    evaluators=[
        LangChainStringEvaluator("qa"),
        LangChainStringEvaluator("cot_qa")
    ]
)

Advanced tracing

Tracing context

from langsmith import tracing_context

with tracing_context(
    project_name="experiment-1",
    tags=["production", "v2"],
    metadata={"version": "2.0"}
):
    # All traceable calls inherit context
    result = my_function()

Manual runs

from langsmith import trace

with trace(
    name="custom_operation",
    run_type="tool",
    inputs={"query": "test"}
) as run:
    result = do_something()
    run.end(outputs={"result": result})

Process inputs/outputs

def sanitize_inputs(inputs: dict) -> dict:
    if "password" in inputs:
        inputs["password"] = "***"
    return inputs

@traceable(process_inputs=sanitize_inputs)
def login(username: str, password: str):
    return authenticate(username, password)

Sampling

import os
os.environ["LANGSMITH_TRACING_SAMPLING_RATE"] = "0.1"  # 10% sampling

LangChain integration

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

# Tracing enabled automatically with LANGSMITH_TRACING=true
llm = ChatOpenAI(model="gpt-4o")
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant."),
    ("user", "{input}")
])

chain = prompt | llm

# All chain runs traced automatically
response = chain.invoke({"input": "Hello!"})

Production monitoring

Hub prompts

from langsmith import Client

client = Client()

# Pull prompt from hub
prompt = client.pull_prompt("my-org/qa-prompt")

# Use in application
result = prompt.invoke({"question": "What is AI?"})

Async client

from langsmith import AsyncClient

async def main():
    client = AsyncClient()

    runs = []
    async for run in client.list_runs(project_name="my-project"):
        runs.append(run)

    return runs

Feedback collection

from langsmith import Client

client = Client()

# Collect user feedback
def record_feedback(run_id: str, user_rating: int, comment: str = None):
    client.create_feedback(
        run_id=run_id,
        key="user_rating",
        score=user_rating / 5.0,  # Normalize to 0-1
        comment=comment
    )

# In your application
record_feedback(run_id="...", user_rating=4, comment="Helpful response")

Testing integration

Pytest integration

from langsmith import test

@test
def test_qa_accuracy():
    result = my_qa_function("What is Python?")
    assert "programming" in result.lower()

Evaluation in CI/CD

from langsmith import evaluate

def run_evaluation():
    results = evaluate(
        my_model,
        data="regression-test-set",
        evaluators=[accuracy_evaluator]
    )

    # Fail CI if accuracy drops
    assert results.aggregate_metrics["accuracy"] >= 0.9, \
        f"Accuracy {results.aggregate_metrics['accuracy']} below threshold"

Best practices

1. Structured naming - Use consistent project/run naming conventions
2. Add metadata - Include version, environment, user info
3. Sample in production - Use sampling rate to control volume
4. Create datasets - Build test sets from interesting production cases
5. Automate evaluation - Run evaluations in CI/CD pipelines
6. Monitor costs - Track token usage and latency trends

Common issues

Traces not appearing:

import os
# Ensure tracing is enabled
os.environ["LANGSMITH_TRACING"] = "true"
os.environ["LANGSMITH_API_KEY"] = "your-key"

# Verify connection
from langsmith import Client
client = Client()
print(client.list_projects())  # Should work

High latency from tracing:

# Enable background batching (default)
from langsmith import Client
client = Client(auto_batch_tracing=True)

# Or use sampling
os.environ["LANGSMITH_TRACING_SAMPLING_RATE"] = "0.1"

Large payloads:

# Hide sensitive/large fields
@traceable(
    process_inputs=lambda x: {k: v for k, v in x.items() if k != "large_field"}
)
def my_function(data):
    pass

References

• Advanced Usage - Custom evaluators, distributed tracing, hub prompts
• Troubleshooting - Common issues, debugging, performance

Resources

• Documentation: https://docs.smith.langchain.com
• Python SDK: https://github.com/langchain-ai/langsmith-sdk
• Web App: https://smith.langchain.com
• Version: 0.2.0+
• License: MIT

Agent operating procedure

1. Check the environment. Confirm hardware, framework and server versions, model format, and expected load.
2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
3. Run a small version first. Serve or log a single request or run end to end before scaling.
4. Execute the full task using the instructions and references above.
5. Validate the result. Measure latency, throughput and output correctness against a reference; check resource usage and costs.
6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.

If this happens	Do this
The server fails to start or OOMs	Check model size versus memory, quantization and parallelism settings.
A function, flag or endpoint in these instructions is missing in the installed version	Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous	Ask the user, or state the assumption explicitly before running.

Integrity rules

• Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
• Do not expose services or credentials publicly; confirm cloud costs before provisioning.
• Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
• Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

• phoenix-observability: Open-source AI observability platform for LLM tracing, evaluation, and monitoring.
• llamaindex: Data framework for building LLM applications with RAG.
• langchain: Framework for building LLM-powered applications with agents, chains, and RAG.