Langsmith Testing

📊 LangSmith Testing SKILL

Purpose

CRITICAL for RAG: Every RAG query MUST be traced in LangSmith for observability. Silent failures (bad retrieval, hallucinations) are only detectable through tracing.

Auto-Trigger Conditions

Activate when:

• User mentions: "LangSmith", "trace", "evaluation", "tracking"
• RAG query execution
• Evaluation tasks
• Performance debugging

LangSmith Setup

1. Installation

# Python
pip install langsmith langchain

# Environment variables
export LANGSMITH_API_KEY="your-api-key"
export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_PROJECT="rag-demo"

2. Basic Integration

from langchain.callbacks import LangChainTracer
from langsmith import Client

# Initialize client
client = Client(api_key=os.environ["LANGSMITH_API_KEY"])

# Create tracer
tracer = LangChainTracer(project_name="rag-demo")

# Use in chain
chain.invoke(query, config={"callbacks": [tracer]})

Trace Collection

1. Automatic Tracing (Recommended)

from langchain_core.tracers.context import tracing_v2_enabled

with tracing_v2_enabled(project_name="rag-demo"):
    # All operations automatically traced
    result = rag_chain.invoke(query)

2. Manual Tracing

from langchain.callbacks import LangChainTracer

tracer = LangChainTracer(
    project_name="rag-demo",
    example_id="example-123"  # Optional: link to dataset
)

result = chain.invoke(query, config={"callbacks": [tracer]})

3. Trace Metadata

from langchain_core.tracers.context import tracing_v2_enabled

with tracing_v2_enabled(
    project_name="rag-demo",
    metadata={
        "user_id": "user-123",
        "session_id": "session-456",
        "environment": "production"
    }
):
    result = chain.invoke(query)

Metrics Collection

1. Core RAG Metrics

Collected automatically:

• Latency: Total pipeline time (ms)
• Token usage: Input + output tokens
• Steps: Number of LLM calls
• Errors: Failed operations

Example trace data:

{
  "run_id": "abc-123",
  "name": "RAGChain",
  "latency_ms": 1250,
  "total_tokens": 850,
  "prompt_tokens": 600,
  "completion_tokens": 250,
  "steps": [
    {"name": "embedder", "latency_ms": 50},
    {"name": "retriever", "latency_ms": 200},
    {"name": "generator", "latency_ms": 1000}
  ]
}

2. Custom Metrics (Faithfulness, Relevance)

from langsmith import Client

client = Client()

# After getting RAG result
client.create_feedback(
    run_id=run_id,
    key="faithfulness",
    score=0.85,  # 0-1 scale
    comment="Answer is mostly accurate to context"
)

client.create_feedback(
    run_id=run_id,
    key="relevance",
    score=0.92,  # 0-1 scale
    comment="Retrieved docs are highly relevant"
)

See: rag-accuracy-SKILL.md for metric calculations

Trace Validation

1. Query Trace

from langsmith import Client

client = Client()

# Get specific run
run = client.read_run(run_id="abc-123")

# Validate trace
assert run.latency_ms < 2000, "Query too slow"
assert run.error is None, "Query failed"
assert len(run.child_runs) >= 3, "Missing pipeline steps"

2. Batch Validation

from langsmith import Client
from datetime import datetime, timedelta

client = Client()

# Get recent runs
runs = client.list_runs(
    project_name="rag-demo",
    start_time=datetime.now() - timedelta(hours=1)
)

# Aggregate metrics
avg_latency = sum(r.latency_ms for r in runs) / len(runs)
error_rate = sum(1 for r in runs if r.error) / len(runs)

print(f"Avg latency: {avg_latency}ms")
print(f"Error rate: {error_rate:.2%}")

Run Comparisons (A/B Testing)

1. Compare Prompt Variants

from langsmith import Client

client = Client()

# Run A: Original prompt
with tracing_v2_enabled(
    project_name="rag-demo",
    metadata={"variant": "prompt-v1"}
):
    result_a = chain_v1.invoke(query)

# Run B: New prompt
with tracing_v2_enabled(
    project_name="rag-demo",
    metadata={"variant": "prompt-v2"}
):
    result_b = chain_v2.invoke(query)

# Compare in LangSmith UI
# Filter by metadata.variant to see performance difference

2. Automated Comparison

from langsmith import Client

client = Client()

# Get runs for each variant
runs_v1 = client.list_runs(
    project_name="rag-demo",
    filter='metadata.variant == "prompt-v1"'
)

runs_v2 = client.list_runs(
    project_name="rag-demo",
    filter='metadata.variant == "prompt-v2"'
)

# Compare metrics
v1_latency = sum(r.latency_ms for r in runs_v1) / len(runs_v1)
v2_latency = sum(r.latency_ms for r in runs_v2) / len(runs_v2)

print(f"V1 avg latency: {v1_latency}ms")
print(f"V2 avg latency: {v2_latency}ms")
print(f"Improvement: {((v1_latency - v2_latency) / v1_latency) * 100:.1f}%")

Dataset Evaluation

1. Create Dataset

from langsmith import Client

client = Client()

# Create evaluation dataset
dataset = client.create_dataset("rag-eval-v1")

# Add examples
client.create_examples(
    dataset_id=dataset.id,
    inputs=[
        {"query": "What is RAG?"},
        {"query": "How does vector search work?"}
    ],
    outputs=[
        {"expected_answer": "RAG is Retrieval Augmented Generation..."},
        {"expected_answer": "Vector search uses embeddings..."}
    ]
)

2. Run Evaluation

from langsmith import Client
from langsmith.evaluation import evaluate

client = Client()

# Define evaluator
def faithfulness_evaluator(run, example):
    # Calculate faithfulness score
    score = calculate_faithfulness(
        run.outputs["answer"],
        run.outputs["source_documents"]
    )
    return {"score": score}

# Run evaluation
results = evaluate(
    lambda inputs: rag_chain.invoke(inputs["query"]),
    data="rag-eval-v1",
    evaluators=[faithfulness_evaluator],
    project_name="rag-eval-results"
)

print(f"Avg faithfulness: {results['aggregate']['faithfulness']['mean']}")

Error Detection

1. Failed Retrievals

from langsmith import Client

client = Client()

# Find runs with no retrieved documents
failed_runs = client.list_runs(
    project_name="rag-demo",
    filter='outputs.source_documents.length == 0'
)

for run in failed_runs:
    print(f"Query: {run.inputs['query']}")
    print(f"Reason: No relevant documents found")

2. High Latency Queries

from langsmith import Client

client = Client()

# Find slow queries (>2s)
slow_runs = client.list_runs(
    project_name="rag-demo",
    filter='latency_ms > 2000'
)

for run in slow_runs:
    print(f"Query: {run.inputs['query']}")
    print(f"Latency: {run.latency_ms}ms")

    # Identify bottleneck
    for child in run.child_runs:
        if child.latency_ms > 1000:
            print(f"  Bottleneck: {child.name} ({child.latency_ms}ms)")

Best Practices

1. Always Trace Production

# Enable tracing in production
import os

os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_PROJECT"] = "rag-production"

# Sample rate for high-volume apps (optional)
os.environ["LANGCHAIN_TRACING_SAMPLING_RATE"] = "0.1"  # 10% of requests

2. Use Meaningful Project Names

# Bad: Generic names
project_name = "test"

# Good: Environment + purpose
project_name = f"rag-{environment}-{feature}"  # "rag-prod-hybrid-search"

3. Add Context with Metadata

with tracing_v2_enabled(
    project_name="rag-demo",
    metadata={
        "user_tier": "premium",
        "retriever_type": "hybrid",
        "llm_model": "claude-3-sonnet",
        "chunk_size": 512
    }
):
    result = chain.invoke(query)

4. Monitor Key Metrics

Daily checks:

• Avg latency < 2000ms
• Error rate < 1%
• Faithfulness score > 0.7
• Relevance score > 0.7

Weekly checks:

• Compare A/B test variants
• Review slow queries
• Identify common failure patterns

Integration with Other Skills

Workflow

User Query
   ↓
LangSmith Tracing (THIS SKILL)
   ↓
RAG Pipeline Execution
   ↓
Trace Collection (automatic)
   ↓
Metrics Calculation (rag-accuracy-SKILL.md)
   ↓
Feedback to LangSmith
   ↓
Dashboard Analysis

Related Files

Troubleshooting

Issue: Traces not appearing

Check:

1. LANGCHAIN_TRACING_V2=true set?
2. LANGSMITH_API_KEY valid?
3. Project name correct?

Solution:

import os
print(os.environ.get("LANGCHAIN_TRACING_V2"))  # Should be "true"
print(os.environ.get("LANGSMITH_API_KEY"))     # Should be set

Issue: High latency in traces

Check:

1. Which step is slow? (embedder, retriever, generator)
2. Network latency to LangSmith?

Solution:

# Disable tracing for testing
os.environ["LANGCHAIN_TRACING_V2"] = "false"
# If latency drops → LangSmith network issue
# If latency same → RAG pipeline issue

Example: Complete RAG with LangSmith

import os
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.chains import RetrievalQA
from langchain_core.tracers.context import tracing_v2_enabled
from langsmith import Client

# Setup
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_PROJECT"] = "rag-demo"

# Build RAG chain
embeddings = OpenAIEmbeddings()
vector_store = FAISS.load_local("./index", embeddings)
retriever = vector_store.as_retriever(search_kwargs={"k": 5})

llm = ChatOpenAI(model="gpt-4", temperature=0)
chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    return_source_documents=True
)

# Query with tracing
query = "What is retrieval augmented generation?"

with tracing_v2_enabled(
    project_name="rag-demo",
    metadata={"environment": "demo"}
) as cb:
    result = chain.invoke(query)
    run_id = cb.run_id

# Add feedback
client = Client()
client.create_feedback(
    run_id=run_id,
    key="faithfulness",
    score=0.9,
    comment="Accurate answer"
)

print(f"Answer: {result['result']}")
print(f"Trace: https://smith.langchain.com/public/{run_id}")

Last Updated: 2025-12-04 Version: 1.0 Priority: CRITICAL (Auto-loads for all RAG queries)