Traces shows the step-by-step execution behind a request — the retrieval, tool calls, and model steps — so you can see exactly how an answer was produced.

What a Trace Is

A trace is one user message the agent processed, made up of spans — the individual recorded steps inside that turn. The most common spans are LLM (a model call), Tool (a tool invocation), Retrieval (fetching knowledge chunks), Reranker (re-ordering retrieved chunks) and Embedding (turning text into vectors). You may also see steps such as Semantic Search, Topic Selection and guardrail checks.

The Traces Table

Each row is one processed user message: Input is that message, Spans counts the recorded steps, Tokens sums input and output tokens across those steps, Cost is the model cost of that turn, and Latency is how long the user waited for the answer. The table lists Trace ID, Input, Start Time, Latency, Tokens, Cost, and Spans, with a stats strip above summarizing Total Traces, Avg Spans per Trace, Avg Latency, and Avg Tokens Consumed. To see the agent's reply, open the trace.

Inspecting a Trace

Click a trace to open its timeline, which walks through the steps the agent took — the model calls, the tools it invoked, and the knowledge it retrieved, with timing for each. This is the place to diagnose a bad answer: walk the steps to see what the agent retrieved, which tool it called, and where the response went wrong. Filter by date range or attribute, and sort as needed. The timeline doesn't auto-refresh once opened; use the refresh button beside the Trace and Metric tabs to re-read it.