Skip to main content
TruLayer models an AI application as a hierarchy of traces (end-to-end units of work) made up of spans (individual steps), grouped optionally into sessions (user conversations). Everything else — evals, feedback, metrics — attaches to those traces.

​
The hierarchy

Session
 ├── Trace (one user request, one cron run, one agent step)
 │    ├── Span (LLM call)
 │    ├── Span (vector search)
 │    ├── Span (tool call)
 │    └── Span (another LLM call)
 └── Trace (the next request in the same conversation)

Attachments on any trace:
 ├── Event       (discrete log within a trace)
 ├── Feedback    (human label — thumbs up/down, score, comment)
 └── Eval result (automated score — correctness, hallucination, etc.)

​
Concepts

Trace

One end-to-end unit of work — typically one user request or one agent turn.

Span

A step within a trace — an LLM call, retrieval, tool call, or custom code block.

Session

A series of traces that belong to the same user conversation or workflow.

Eval

An automated score for a trace — correctness, hallucination, safety, etc.

Feedback

A human label attached to a trace — thumbs up/down, score, comment.

Metric

An aggregated view across traces — error rate, p95 latency, cost.

​
How it all fits together

You instrument your app with an SDK. Every LLM call, retrieval, or tool call becomes a span inside a trace. Traces are batched and shipped to the TruLayer ingest API. In the background:
  • Evals run against traces and produce scores (stored as eval results).
  • Feedback is attached to traces directly from your app or through the dashboard.
  • Metrics are computed on a rolling window and surfaced in the dashboard and via the API.
  • Failures are clustered and routed to alerting channels you configure.
From the developer’s side: all you need to do is instrument. The rest happens automatically.