What it does
Langfuse traces, evaluates, and monitors LLM applications. It brings prompts, model calls, retrieval steps, and agent actions into a shared view, with self-hosting available.
Why it’s interesting
A slow or incorrect AI answer can involve several steps. Following a trace helps you find which retrieval or model call needs attention, instead of guessing from the final response.
Real-world use case
Trace a support assistant and inspect the retrieved context when it answers incorrectly. Compare a revised prompt against saved examples before shipping the change.
What I can learn
- 01
Distributed tracing
follow nested operations through an AI workflow.
- 02
Evaluation design
turn real examples into repeatable quality checks.
- 03
Prompt versioning
connect application behavior to the prompt that produced it.
One thing to explore
Shared data models
Follow one observation type through the shared package and note how it relates to a trace. Sketch the relationship between an application request and its model calls.
Explore the source (opens in a new tab)10-Minute Challenge
Sketch the trace for a RAG request
Before you start: A browser and the repository README. No API key or local installation needed.
- Read the observability section in the README.
- Draw a trace with three steps: retrieve documents, build a prompt, and call the model.
- Write down which inputs, outputs, and timings you would capture to debug a bad answer.