Traces
Browse and inspect recorded traces.
Overview
The Traces view lists every trace recorded by TraceLLM. Each row shows the trace ID, timestamp, status, model, latency, token count, and step count. Click any trace to inspect its full execution graph in the detail panel.
The view fetches data from GET /traces, which supports filtering by status, model, project, environment, latency range, and token count range.
Filtering Traces
The API supports the following query parameters to narrow down trace results:
| Parameter | Type | Description |
|---|---|---|
| status | string | Filter by status: success, warning, or failed |
| model | string | Filter by model name (e.g. gpt-4o) |
| project_id | string | Filter by project identifier |
| environment | string | Filter by environment: development, staging, production |
| latency_min | float | Minimum latency in milliseconds |
| latency_max | float | Maximum latency in milliseconds |
| token_min | int | Minimum token count |
| token_max | int | Maximum token count |
| limit | int | Max results to return (1-200, default 50) |
Trace Inspector
Clicking a trace opens the inspector, which displays the full trace document including all metadata and every captured step. The inspector fetches a single trace via GET /traces/{trace_id}.
The inspector shows:
- Trace ID, prompt, and response text
- Model name, project, and environment labels
- Total latency, token count, retry count, and status
- Complete step list with timing, inputs, outputs, and success status per step
- Failure reason when status is failed or warning
Step Details
Each step in a trace contains structured data about a single atomic operation:
| Field | Type | Description |
|---|---|---|
| step_id | string | UUID4 unique identifier for this step |
| tool_name | string | Name of the tool or operation (e.g. vector_retrieval) |
| input | object | Input parameters passed to the tool |
| output | object | Output or result returned by the tool |
| duration | float | Wall-clock time in milliseconds |
| success | boolean | Whether the step completed without error |
| timestamp | datetime | ISO 8601 timestamp of execution |
Export from Dashboard
Individual trace data can be copied from the inspector panel. For bulk export, use the tracellm export CLI command which supports JSON and CSV formats.
Tip
tracellm replay {trace_id} command to replay any trace from the terminal — no dashboard needed.