#Production & LangSmith

Shipping LangChain apps needs observability, evals, security, and deployment. LangSmith is the official tracing and evaluation platform.

#LangSmith

• Tracing: LLM calls, tools, latency, tokens per run
• Datasets & evals: regression suites, automated scoring
• Monitoring: production anomaly detection (Engine evolves over time)

Site: smith.langchain.com

#Enable tracing

export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_API_KEY="lsv2_pt_..."
export LANGCHAIN_PROJECT="my-app-prod"

Often zero code changes. Inspect full message traces and tool I/O in the UI.

#Evaluations

1. Create a dataset (inputs + reference outputs)
2. Batch-run your agent
3. Score with LLM-as-judge or rules

For RAG: faithfulness, answer relevance, context precision.

#Dev vs prod

#Security

Prompt injection via tool outputs, least-privilege tools, PII middleware, step/model cost controls.

#Deployment

• Self-hosted FastAPI + Postgres checkpointer
• LangGraph Platform (managed)
• Queue workers for long jobs

#Pin dependencies

langchain>=1.0,<2
langgraph>=1.0,<2

Run smoke tests and LangSmith eval gates in CI.

#Next steps

• Practical Examples
• Resources