Langfuse
Langfuse is an open-source platform for observability and evaluation of AI applications. It captures hierarchical traces of every model call, tool call and retrieval step including cost and latency, and adds prompt management, test datasets, experiments and evaluations.
Langfuse is an open-source platform for tracing and evaluating AI applications: nested traces per model and tool call, prompt management, datasets and evaluations. Self-hostable, including in the EU.
Explanation
Without traces an AI application is a black box surrounded by opinions about whether it got better. Langfuse turns that into data: for every request you see which passages were retrieved, which tools were called and how many tokens were consumed — filterable by user, session, cost and duration. Evaluations build on top, either heuristic, model-as-judge or human review. The practical value shows when switching models: you compare on the same test dataset instead of deciding by feel. Because the platform is fully open source and built on open components, it can run in your own network and therefore in the EU — the usual route when conversation content is personal data. Some add-on features require a licence key.
Marketing Relevance
For decision-makers Langfuse is the tool that makes AI application quality acceptable: error rate on a fixed test set, cost per request, duration per step. That turns "feels good" into a number you can record in an acceptance report.
Example
Before switching the chat model, 80 typical visitor questions are stored as a test dataset. Both models run against it and Langfuse shows hit rate, cost and duration per question. The decision follows the differences, not an impression.
Common Pitfalls
Traces quickly contain personal content — without masking and a retention period you create a second data store of conversation data. And self-hosting brings real effort, because database, analytics store and object storage must be operated too.
Origin & History
Langfuse launched in 2023 as an open-source tracing tool for LLM applications and grew into a broader platform as demand for solid evaluations rose. It supports OpenTelemetry and numerous frameworks; since January 2026 the company has been part of ClickHouse.
Further Resources
Marketing Use Cases
Traceability of chat and agent runs down to tool level.
Evaluations as acceptance criteria before a model switch.
Cost control per use case and per user group.
Versioned prompts with documented changes.
Frequently Asked Questions
Can I run Langfuse in the EU?
Yes. The platform is open source and fully self-hostable, for instance in an EU data centre. That is the usual route when conversation content contains personal data.
Langfuse or LangSmith?
Langfuse is open source and self-hostable, LangSmith a commercial service most tightly integrated with the LangChain ecosystem. Those prioritising data sovereignty usually pick Langfuse.
What belongs in a test dataset?
Real, typical requests with the expected core of the answer — including the awkward cases. A dataset of easy questions only confirms what already works.
Related Services
Go deeper: Agentic AI Hub · Model comparison 2026