KNet Traces

KNet Traces records what your LLM and agent applications actually do at runtime, so you can debug a bad answer, find the slow step, and track what it all costs.

It lives at flow.knetai.com/traces and shares your KNet organisation and workspace, so a project in Traces is the same project you already work in elsewhere on the platform.

Why you want it

A single user request to an agent can fan out into dozens of model calls, retrievals, and tool invocations. When the answer comes back wrong, slow, or expensive, the logs tell you almost nothing: you get the final output and no record of how it was reached.

Traces keeps the whole tree. You can open any request and read every step in order, with the inputs, outputs, latency, and token cost attached to each one.

Core concepts

Trace — one end-to-end request through your application. A user question, a scheduled job, a webhook. This is the unit you search for and share.

Observation — a single step inside a trace: a model call, a retrieval, a tool invocation, a nested span. Observations form a tree, so you can see which step called which.

Session — a group of traces belonging to one conversation or user journey, so a multi-turn chat reads as one thread rather than scattered requests.

Score — a rating attached to a trace, observation, or session. Scores can come from a human reviewer, an automated evaluator, or your own code, and are what let you say whether quality is moving in the right direction.

What you can do with it

• Tracing — inspect the full step tree for any request, with inputs, outputs, latency, and cost per step.
• Sessions and Users — follow a conversation across turns, or see everything one user did.
• Prompts — manage prompt versions outside your deployment, so changing a prompt is not a code release.
• Playground — take a captured call, edit the prompt or parameters, and re-run it against the real model.
• Scores, Evaluators and Human Annotation — attach quality ratings automatically or by hand, and review them in one queue.
• Datasets and Experiments — collect real traffic into a test set and compare versions of your application against it.
• Dashboards and Widgets — track cost, latency, volume, and quality over time, with dashboards you can build yourself.
• Alerts — get told when something moves in the wrong direction instead of discovering it later.

Projects and access

Traces uses the KNet organisation and workspace you are already a member of. A workspace maps to a project in Traces, and your access is inherited — there is no separate account to create and no second set of permissions to keep in sync.

Next steps

Ready to connect an application? See Send your first trace.

Getting help

If something here does not answer your question, see Getting help. A person reads every support request.