Skip to main content
A trace records a single request end to end. For a multi-step workflow, child calls are nested under one trace so you can follow the whole chain. See Distributed tracing for how traces are assembled across the gateway and your own spans.

​
Open the Trace Explorer

Traces are scoped to a project. In the dashboard, select a project and open Traces — the Trace Explorer. It opens on the most recent traces in the project, newest first.

​
The trace list

Each row is one trace. The list shows:
ColumnWhat it shows
TimeWhen the request started. Sorted newest-first by default.
StatusOK for a successful request, ERROR for a failed or timed-out one.
DurationWall-clock time for the whole trace, in milliseconds.
Provider / ModelThe provider and the model the request resolved to.
In Tok / Out TokInput (prompt) and output (completion) token counts.
CostEstimated cost in USD, or — when no cost is available.
SessionThe session the request belongs to, when one is set.
SpansHow many spans make up the trace — one for a single call, more for a multi-step workflow.
TimelineA bar showing the trace’s duration relative to others on the page.
Trace IDThe trace’s unique ID, with a button to copy it.
Click a column header to sort by Time, Duration, In Tok, Out Tok, Provider / Model, or Cost. Use the rows-per-page control to show 25, 50, or 100 traces at a time and page through the results.

​
Filter

The toolbar above the list narrows what you see. Filters combine — a trace must match every active one.
  • Time range — a preset window from the last 10 minutes through the last 60 days (the list defaults to the last 8 hours), or an absolute start and end time.
  • Status — All, OK, or Errors (errors include timeouts).
  • Provider — one or more providers.
  • Model — one or more models.
  • Tool Calls — Any Tool Calls, With (traces that made tool calls), or Without.
  • Tokens — a minimum and/or maximum total token count.
Use Clear all filters to reset every filter at once. When a query returns nothing, the Trace Explorer tells you whether the project has no traces in the window or whether your filters excluded them all, and offers to clear the filters or widen the time range to the last 7 days.

​
Open a trace

Click any row to open the trace. The header summarizes the whole trace — status, duration, span count, total tokens, cost, and counts of LLM calls and tool calls — and lets you download the full trace as JSON. Below the header, the trace is laid out as a waterfall: every span on its own row, nested under its parent and positioned by when it ran, so you can see the shape of a multi-step workflow and where the time went. Spans are color-coded by kind — application, LLM call, embedding, retrieval, and tool — and failed spans stand out. Select a span to open its detail panel, which has tabs for:
  • Overview — the span’s timing and attributes.
  • Input — the prompt or input sent.
  • Output — the completion or result returned.
  • Tool Calls — any tool calls the span made.
  • Events — timestamped events recorded on the span.
  • Raw — the span’s full underlying data.
The Input and Output tabs show prompt and completion text only when content capture is enabled — it is on by default. For the full list of fields a span can carry, see Span attributes.

​
Access to traces

Who can view traces — and production traces specifically — is governed by permissions. See Roles & permissions.

​
Next steps

Span attributes

What each span records.

Distributed tracing

How calls are grouped under one trace.

Content capture

Control whether prompt and completion text is recorded.

Roles & permissions

Control who can see traces.