Traces

Traces is the raw evidence layer for a project. Use it when you need to inspect the exact production requests behind a metric, Behavior, evaluation failure, or Improve diagnosis. Each trace represents an end-to-end request. A trace can include model spans, tool calls, retrieval, guardrails, orchestration, and custom spans sent through the API, SDK, or proxy.

​

What the table shows

The Traces table includes:

• Started At - when the request began.
• Name - the trace name or workflow name.
• Latency - total request time.
• Cost - total provider cost when available.
• Tokens - input, output, and total token counts.
• First span input and Last span output - quick context without opening the trace.
• Tags and Attributes - metadata attached during instrumentation.
• IDs - trace and span identifiers for debugging and API workflows.

You can choose visible columns from the table controls. The app keeps a prompt filter available for workflows that need to narrow traces to a particular prompt without turning the table into prompt-only aggregates.

​

Open a trace

Open a trace when you need to inspect the full execution. Use the trace viewer to review:

• Span hierarchy and timing.
• Model input and output.
• Tool calls and tool responses.
• Retrieval and search context.
• Tags, attributes, and IDs.
• Evaluator scores attached to spans.
• Prompt and deployment context when available.
• Provider, model, cost, token, and latency details for model spans.

​

Trace metadata model

Tags and attributes are the bridge between your application and Adaline workflows: Consistent metadata makes filters, exports, Monitor segmentation, Behavior analysis, and dataset creation much more useful. See Instrumentation metadata for the metadata plan to send from your application.

​

Behavior-aware trace review

When you arrive from a Behavior, Traces can narrow the table to spans that belong to that Behavior. This is useful when a Behavior represents only part of a larger request: you can inspect the full trace while still anchoring the search to the member spans that explain the Behavior.

​

Use traces with the rest of Platform

Traces are connected to every improvement workflow:

• Monitor aggregates traces into charts.
• Behaviors clusters traces into repeated patterns.
• Datasets can store examples from traces.
• Improve uses traces as evidence for prompt candidates.
• Evaluators can score spans for continuous quality monitoring.