Langfuse: LLM Tracing, Evaluations and Retention Costs
What does Langfuse do?
Langfuse is an AI engineering platform covering traces, prompt management, evaluations and metrics. Its documentation describes cloud and self-hosted approaches, with SDK and OpenTelemetry integration options. The practical goal is to connect an application outcome to the retrieval, model and tool steps that produced it.
Who is it for?
Developers maintaining a multi-step AI application can use traces to distinguish retrieval errors from model errors and external-tool failures. A team with only occasional manual prompts may not need a separate tracing platform. It becomes more useful when failures must be reproduced across versions or compared across real workflows.
What are the plan boundaries?
The pricing page lists Hobby as free with 50,000 units per month, 30-day data access and two users. Core is $29 per month with 100,000 included units and 90-day data access; additional usage is metered. A unit is not necessarily one user conversation. Check current unit definitions, retention and feature limits for your instrumentation plan.
How should you test trace usefulness?
Instrument a small synthetic support assistant with a retrieval step and one simulated tool failure. Record a correlation ID, prompt version and expected outcome for each case. Ask a colleague who did not run the trial to find the failing step from the trace alone. A visually rich timeline is less useful if key inputs or error boundaries are missing.
What makes the evaluation actionable?
Use a labeled set containing answerable, ambiguous and unanswerable questions. Keep the evaluator definition versioned alongside the prompts. Inspect disagreement between human judgments and automated scores rather than averaging it away. Separate correctness, source support and response time so an improvement in one metric does not hide a regression in another.
What can go wrong?
Traces can contain private prompts, documents and tool outputs. Plan redaction before ingestion and check who can view the project. Self-hosting adds responsibility for storage, backups and upgrades. Instrumentation also has failure modes: verify that logging problems do not silently alter the application response or leak secrets into an error message.
Which alternatives should you compare?
Compare Braintrust for an evaluation-centered investigation workflow and LangSmith for tracing and evaluation alongside its broader agent platform. Compare using the same synthetic incident, data retention requirement and expected monthly trace volume.
Sources checked October 27, 2026: Official documentation · Pricing and access conditions.