Traces & Observations
The Traces view is your primary debugging surface — every request your app makes, with full inputs, outputs, timing, cost, and the nested steps that produced the result.
Open it from Tracing in the left nav. To get traces flowing first, see Sending Data In.
The trace table
The table lists traces newest-first. Columns include name, timestamp, user, session, latency, total cost, token usage, tags, and any scores. You can:
- Search by trace ID, name, user ID, or tags.
- Filter by timestamp, latency, cost, scores, user, session, tags, level, and environment.
- Sort and customize columns (visibility is remembered per project).
- Select rows for batch actions (export, delete).
The observation tree
Open a trace to see its observation tree — the nested steps that make up the request:
| Observation kind | Represents |
|---|---|
| Span | A unit of work — a function, retrieval step, or tool call |
| Generation | An LLM call, with model, prompt, completion, token counts, cost, and parameters |
| Event | A point-in-time log line |
Each observation has a level (Debug / Default / Warning / Error) so you can filter noise and jump straight to errors. Select any node to inspect its input, output, metadata, and timing.
Detail views
A trace offers several views:
- Details — the observation tree with full input/output inspection.
- Timeline — observations laid out chronologically, so you can see what ran when and where time went.
- Graph — for agent workflows, a node-graph of the execution (see Agent Graphs).
Working with a trace
Inspect
Walk the observation tree, open the slow or failing step, and read its input/output and metadata.
Annotate
Add scores (manual annotation) and comments to a trace or observation to capture quality judgments and notes for your team. See Scores.
Organize
Tag traces, bookmark/star important ones, and optionally publish a trace to share a read-only view.
Export
Export a trace (and its observations) as JSON, or batch-export from the table.
Trace inputs and outputs can contain sensitive data. Govern access with Roles & Access, and use AI Guardrails to detect PII/PCI/PHI in your traffic.
Next steps
- Sessions — group traces into conversations.
- Agent Graphs — visualize agent execution.
- Scores — measure trace quality.