AI applications are notoriously hard to debug. A chain of LLM calls, tool invocations, and conditional logic creates complex execution paths that need visibility.

LangSmith for Tracing

LangSmith captures every step of your chain execution:

• Traces: Full execution tree showing each node, its inputs, outputs, and timing
• Token usage: Per-step and total token counts with cost estimates
• Latency breakdown: Where time is being spent (model inference vs. tool calls)
• Error tracking: Which step failed and why, with full stack traces

Setting Up Tracing

import os
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_API_KEY"] = "your-langsmith-key"
os.environ["LANGCHAIN_PROJECT"] = "my-production-app"

Once set, all LangChain operations are automatically traced. No code changes needed.

Production Patterns

Fallback chains: If the primary model fails, fall back to an alternative:

primary = ChatOpenAI(model="gpt-4o")
fallback = ChatAnthropic(model="claude-sonnet-4-20250514")
robust_model = primary.with_fallbacks([fallback])

Retry logic: Automatically retry on transient failures:

from langchain_core.runnables import RunnableConfig
chain.invoke(input, config=RunnableConfig(max_retries=3))

Streaming: For user-facing applications, stream tokens as they're generated:

for chunk in chain.stream({"question": "Explain quantum computing"}):
    print(chunk, end="", flush=True)

Caching: Avoid redundant LLM calls for identical inputs:

from langchain_core.globals import set_llm_cache
from langchain_community.cache import SQLiteCache
set_llm_cache(SQLiteCache(database_path="cache.db"))

Evaluation with LangSmith

Create evaluation datasets and run your chain against them:

1. Collect representative input-output pairs
2. Define evaluation criteria (correctness, relevance, safety)
3. Run automated evals on every code change
4. Compare performance across chain versions

This creates a CI/CD pipeline for AI quality, catching regressions before they reach production.