Langfuse Overview

Traces allow you to track every LLM call and other relevant logic in your app.

Trace detail of a RAG answer with the generation selected. Hover the cost and token values in the generation header to see the cost and usage breakdowns, expand the collapsed chat history, then select the retrieval span in the tree to see the context it returned.

Sessions allow you to track multi-step conversations or agentic workflows.

Trace detail of one support-chat turn. Click the “session” badge in the trace header to open the session: every turn of the conversation in one view, with the user and assistant messages in order. Click the last turn in the list on the left to jump to it.

Debug latency issues by inspecting the timeline view.

Trace detail of an agent trace. Switch from “Tree” to “Timeline” to see each step as a bar on a time axis with its duration and cost, so the slowest steps stand out. Click one of the llm.chat calls to open it, then switch back to “Tree”.

Add your own userId to monitor costs and usage for each user. Optionally, create a deep link to this view in your systems.

Users list with each user's first and last event, trace count, tokens and cost. Click a user to open their page with total observations, tokens and cost, their traces, and the “Scores” tab with every score on their traces.

LLM agents can be visualized as a graph to illustrate the flow of complex agentic workflows.

Trace detail of an agent trace, “Graph” tab next to Tree and Timeline. In the default Aggregated mode, repeated steps collapse into one node such as “llm.chat (3/3)”; select it to see its observations. Switch the toggle to Expanded to see the run as it happened, click a tool call node, then switch back to Aggregated.

See quality, cost, and latency metrics in the dashboard to monitor your LLM application.

Home dashboard with traces by name, model costs and scores. Change the time range from “Past 1 day” to “Past 7 days”, then scroll through observations by level, model usage and cost, user consumption, score trends and latency percentiles.

Get notified over Slack, GitHub Actions, or Webhooks when a metric crosses a threshold with alerts.

Create a new prompt via UI, SDKs, or API.

Prompts > “New prompt”. Enter the name order-status-reply and a text prompt with the variables {{customer_name}} and {{order_id}}; Langfuse lists the detected variables below the editor. Keep “Set the production label” checked and click “Create prompt”: version 1 opens with the production and latest labels.

Collaboratively version and edit prompts via UI, API, or SDKs.

Prompts > support-reply > “New version”. Edit the system message, add the commit message “Confirm order numbers” and click “Save new prompt version”. The new version appears on top of the version list with its commit message and the latest label.

Deploy prompts to production or any environment via labels - without any code changes.

Prompts > qa-answer-with-context-chat with the latest version (#12) selected. Click “Add label”, tick production under “Promote to production?” and click “Save and promote to production”; the production label moves to #12. The video then selects #9 and promotes it back to production.

Compare latency, cost, and evaluation metrics across different versions of your prompts.

Prompts > qa-answer-with-context-chat > “Metrics” tab: one row per version with median latency, input and output tokens, cost, generation count and score columns such as accuracy and helpfulness. Switch back with the “Versions” tab.

Instantly test your prompts in the playground.

Prompts > movie-critic, version 3. Click “Playground” > “Fresh playground”, fill the variables criticlevel (expert) and movie (Dune: Part Two) in the Variables panel and click “Submit” to get the model's answer in Output.

Link prompts with traces to understand how they perform in the context of your LLM application.

Trace detail with a generation that used movie-critic v3. Click the prompt badge “movie-critic - v3” in the generation header to open the prompt, then open its “Linked Generations” tab and the “Metrics” tab with latency, cost and scores per version.

Track changes to your prompts to understand how they evolve over time.

Prompts > support-reply with the latest version (#5) selected. Hover an older version in the version list and click its compare icon; the “Changes v2 → v5” dialog shows the content and config diff. Close it.

Plot evaluation results in the Langfuse Dashboard.

Scores > “Analytics”. Pick helpfulness as the first score to see its statistics, trend over time and distribution, then pick faithfulness as the second score to compare both: means, matched pairs, correlation and a side-by-side distribution.

Collect feedback from your users. Can be captured in the frontend via our Browser SDK, server-side via the SDKs or API.

Scores page. Search name:user-feedback value:<0.5 to list the thumbs-down feedback the chat widget sent through the API, with the customer's comment such as “Still charged twice.” Click a row's trace to open the conversation in a side panel.

Run fully managed LLM-as-a-judge evaluations on production or development traces. Can be applied to any step within your application for step-wise evaluations.

Evaluators > “New evaluator” > template “Classify Input Topic” (LLM-as-a-judge). Filter the sample observations to the support chatbot, map the {{input}} variable to payload.message under “Map variables to data”, run the judge on one sample to see its score and reasoning, then click “Create evaluator” and confirm the rule in the “Evaluator saved” dialog.

Evaluate prompts and models on datasets directly in the user interface. No custom code is needed.

Datasets > capital-cities > “Run experiment” > “via User Interface” > Configure. Select prompt answer-question, version 3, and model gpt-4.1, keep the dataset capital-cities, skip evaluators, name the run and click “Run Experiment”. A toast confirms the experiment was triggered.

Baseline your evaluation workflow with human annotations via Annotation Queues.

Human Annotation > “New queue”: name it “Error Analysis”, add the “First Pass” score config and click “Create queue”. In Tracing, select three traces, choose “Add to Annotation Queue” in the selection bar and pick “Error Analysis”. Open the queue with “Process queue” and annotate the first item in the Annotate panel.

Add custom evaluation results, supports numeric, boolean and categorical values.

POST /api/public/scores

Add scores via Python or JS SDK.

from langfuse import get_client

langfuse = get_client()

langfuse.create_score(
  trace_id="123",
  name="my_custom_evaluator",
  value=0.5,
)