Trace viewer

The trace viewer at langfuse.nufi.me records every AI request NUFI handles — the prompt, the reply, the latency, the token counts, the cost, and the user identity. This is what you open when a user says "the AI got it wrong" or "where did our budget go this month?"

Sign in

The first user to register on a fresh install becomes the org owner. Subsequent users you invite from Settings → Members.

The trace viewer has its own user model, separate from the chat. This is intentional — it limits who can read other users' conversations to a small set of trusted operators.

What is a trace?

A trace is a top-level event — usually one chat completion. Inside a trace, observations record nested calls (one per AI call, one per tool call, one per safety filter). Each observation has its own duration, model, prompt, reply, and cost.

What you actually do here

Debug a specific user complaint

A user says "the model lied to me at 2 pm".

1. Traces → filter userId = <their id> and timeframe last 2h.
2. Find the trace, click in.
3. The Observations panel shows every step — the system prompt, the user prompt, the assistant reply, any tool calls, any safety filter decisions.

Compare two models on the same prompt

1. Datasets → create a dataset of prompts you care about.
2. Datasets → Run → pick the two models.
3. Read the diff and the latency comparison.

Score a model run

For feedback collection:

1. Open a trace.
2. Click Add score in the observation.
3. Tag it (quality, helpfulness, 1-5, …).

Scores aggregate into the dashboard panels and the Datasets view.

Watch cost

Dashboard → Cost shows daily and monthly spend broken down by model, user, and hardware. Useful for "which team is burning through the budget?" questions.

Retention

By default, traces are kept forever. If your trace store grows too large, ask your operator to set a retention policy — typically 90 days for development, 1 year for production.