What does Langfuse show about one wrong answer?

A trace follows one attempt through the steps your app records, such as fetching a help page and asking a model for an answer. Open the model step to inspect the input and output your app sent to Langfuse. If a step is missing, check whether the app is set up to record it.

For a web app, your developer can connect the server-side AI code to a Langfuse project with its Python or JavaScript library. Match the library version to your installation and keep secret project keys on the server.

Save repeatable test inputs and expected answers in a dataset. Keep the dataset version fixed when you compare two runs. You can add a human score to a recorded result or run an experiment that sends the app's outputs and scores back to Langfuse.

How can you judge two answers before running an experiment?

Write two fictional questions on paper for a help assistant. Give it a sample help page that lists returns within 30 days and says nothing about replacement parts.

Set one rule per question
For “How long do I have to return this?” mark an answer correct if it says 30 days. For “Can I buy replacement parts?” mark it correct if it says the help page gives no answer and offers a way to ask.
Review sample answers
Imagine the first version says 30 days and makes up a replacement-part price. Imagine the second answers both cases correctly. Give each response its own pass or fail and write the reason. The invented price matters even though one answer was correct.
Choose what the trace should explain
For the failed answer, write down which sample page was supplied and which step produced the unsupported price. In a disposable integration, those are the details you would look for in a recorded trace, provided your app sent them.

Which private details should stay out of the trace?

Decide what the trace needs before it leaves your app. Keep secrets, customer messages and private documents out of test traces. Langfuse documents masking for inputs, outputs and metadata, with different coverage depending on the library and how the data is sent.

Verify the actual tracing destination and capture rules before opening a test app. Use only invented data and disable or replace live services with test services. Leave attachments out of this check unless you have verified their separate handling. A masking rule for text fields should not be assumed to cover media. Use the paper example if you cannot confirm the setup.

What do Cloud usage and self-hosting cost?

Langfuse Cloud counts each recorded trace, observation or score as one unit. An observation is one recorded step. One trace with two observations and one score uses four units. Experiments can create billable units too.

On November 12, 2026, Cloud Hobby listed 50,000 units per month. Core listed 100,000 included units for US$29 per month, with usage above the included amount billed in volume tiers. Model usage and any automated judge calls can add separate model-provider costs.

The self-hosted open-source edition uses the MIT license and has no Langfuse usage charge. You still run the application, databases, queue and file storage and handle updates and backups. Self-hosted Enterprise has separate licensing and features.

Does the history window delete old traces?

The Cloud Hobby and Core plans show 30 and 90 days of history, respectively. Older data can remain stored beyond that viewing window. Scheduled deletion is a separate project setting on supported Pro, Enterprise and self-hosted Enterprise plans.

Many tables let you export filtered data as CSV or JSON. Hiding a column in the table does not remove it from the export, so check the file before sharing it. A saved dataset item and its source trace are separate records. Deletion can remove a trace and its scores while the test item remains.