Tracing Structure

You’ve initialized Laminar. Your LLM calls are traced automatically. Now what?

This section covers what to do when auto-instrumentation isn’t enough:

• Trace your own functions — Your agent has logic beyond LLM calls. Wrap functions with @observe to see them in traces.
• Add context — Attach user IDs, session IDs, metadata, and tags so you can filter and debug effectively.
• Control what’s captured — Skip recording inputs/outputs for sensitive operations.

If you just want to trace LLM calls and framework operations, you don’t need this section—Integrations has you covered. Come back here when you want deeper visibility into your own code.

Why Structure Matters

Without structure, each LLM call becomes an isolated trace, making it hard to:

• Understand relationships between calls (routing → tool use → final answer)
• Follow multi-step workflows end-to-end
• Track conversations across turns
• Find the right trace quickly (by user/session/metadata)

Well-structured traces show a clear tree of spans for each request, plus stable identifiers that help you search and group related work.

Quickstart

Start by creating a parent span for your request/turn, then set user ID, session ID, and metadata inside that span so everything downstream inherits it.

TypeScript

import { Laminar, observe } from '@lmnr-ai/lmnr';

Laminar.initialize({
  projectApiKey: process.env.LMNR_PROJECT_API_KEY,
});

export async function handleRequest(userId: string, requestId: string) {
  return observe({ name: 'handleRequest' }, async () => {
    Laminar.setTraceUserId(userId);
    Laminar.setTraceSessionId(`session-${requestId}`);
    Laminar.setTraceMetadata({ environment: process.env.NODE_ENV });

    // ... your LLM calls / tools / business logic ...
  });
}

Python

from lmnr import Laminar, observe
import os

Laminar.initialize()

@observe()
def handle_request(user_id: str, request_id: str):
    Laminar.set_trace_user_id(user_id)
    Laminar.set_trace_session_id(f"session-{request_id}")
    Laminar.set_trace_metadata({"environment": os.getenv("ENVIRONMENT")})

    # ... your LLM calls / tools / business logic ...

Quick Reference

I want to…	Use
Trace a function I wrote	Trace Functions
Trace code that isn’t a function	Trace Parts of Your Code
Make a function show up as a tool call in transcript view	Span Types
Group traces by conversation/workflow	Sessions
Associate traces with users	User ID
Add key-value context to traces	Metadata
Add categorical labels to spans	Tags
Trace images sent to LLMs	Tracing Images
Continue traces across services	Continuing Traces
Track costs for custom LLM spans	LLM Cost Tracking
Ensure spans send in serverless/CLI	Flushing & Shutting Down