Langfuse · Lesson 0001
What Langfuse is — and why “Lang” is a trap
Four popular tools start with “Lang”: LangChain, LangGraph, LangSmith, Langfuse.
People assume they are versions of the same thing. They are not. Split them into two layers and the
whole category becomes obvious.
Two layers, four tools
Build layer — runs your app
- LangChain — assemble LLM apps: prompts, tools, retrieval, chains.
- LangGraph — build agents as stateful graphs (loops, memory).
These are frameworks. They are the code that produces the LLM calls.
Observe layer — watches your app
- LangSmith — trace/eval, native to LangChain/LangGraph. Closed source.
- Langfuse — trace/eval, framework-agnostic. Open source (MIT core).
These are platforms. They sit beside your app and record what it did.
The single sentence: Langfuse competes with LangSmith, not with LangChain/LangGraph.
You can build an app with LangGraph and observe it with Langfuse — they are different layers.
So what is Langfuse, concretely?
An open-source LLM engineering platform (MIT core, YC W23, ~30k GitHub stars,
now owned by ClickHouse). You instrument your app; it records every LLM call, tool call and retrieval
step as a trace, then gives you:
- Observability — cost, latency, full I/O per trace. Debug “why did this cost $4?”
- Evaluations — LLM-as-a-judge, code checks, human annotation, datasets & experiments.
- Prompt management — version, deploy, roll back prompts without a deploy.
- Playground — iterate prompts on real production inputs.
It is OpenTelemetry-native. That matters a lot for us: your Go services
(via the shared core library) already emit OTel. Langfuse is simply an OTel backend — so for
much of your stack, pointing the existing exporter at Langfuse is most of the integration. We go deep
on the Go path in the next lesson.
The win for this lesson: you can place all four “Lang” tools into
Build vs Observe, and you can state in one line why Langfuse is not a rival to LangChain.
If someone says “we use LangChain, we don’t need Langfuse,” that sentence confuses the layers — you now know why.
Check yourself (cover the answers, recall first)
Quick recall
Q1. Which tool builds agents as graphs?
A1. LangGraph builds agents as graphs.
Q2. Which tool traces apps and is open source?
A2. Langfuse traces apps and is open source.
Q3. Which layer does LangSmith belong in?
A3. LangSmith belongs in the observe layer.
Answers are deliberately the same length — no formatting clues.