Observability
Every request your AI application makes is captured as a trace, broken down into nested observations, and enriched with cost, latency, token usage and scores. You can debug a single request, follow a multi-turn conversation, or analyze usage per person.
ℹ
New here? Start with Sending Data In to instrument your application, then come back to explore your traces.
Where to find it
| Page | Sidebar location |
|---|---|
| Traces & Observations | Discover AI > Tracing |
| Sessions | Discover AI > Sessions |
| Users | Gauge AI > Users |
Cost, latency and usage dashboards built on the same data are under Gauge AI. See FinOps.
The data model
| Concept | What it is |
|---|---|
| Trace | One end-to-end request or run of your application, such as one chat turn or one agent task. |
| Observation | A step inside a trace. There are three types: span (work such as a tool call or retrieval), generation (an LLM call with model, tokens and cost) and event (a point in time). Observations nest to form a tree. |
| Session | A group of traces that share a session ID, such as a whole conversation. |
| User | The end user ID attached to traces, for per-user analytics. |
| Score | A quality value on a trace, observation or session, from the API, an evaluator or human annotation. |
text
Session
└─ Trace (one request)
├─ Span: "retrieve context"
│ └─ Generation: "embed query" (model, tokens, cost)
├─ Generation: "answer" (model, tokens, cost, latency)
└─ Event: "guardrail checked"
+ Scores: quality=0.9, helpfulness="good"
Who can see traces
Access is controlled by role. See Members & Roles.
- Listing traces and sessions needs the Tracing & Sessions permission in the Governance module. Among built-in roles, Owner, Department Admin, Security & Compliance and Auditor have it.
- Opening a trace or session to read its input, output and observation content also needs the PII / raw-prompt read permission. Among built-in roles, Owner and Security & Compliance have it. Other roles can list traces but not open them, unless the trace has been shared publicly.
- Users (per-user analytics) needs the Users permission in the Gauge AI module. Among built-in roles, Owner, Department Admin, Security & Compliance and Cost Analyst have it.
Owners and Security & Compliance members can also turn off Trace input and output for the organization or for a project. Prompts and replies are then hidden everywhere in the UI. See Data Privacy.
In this section
Related
- Core Concepts: Tracing and Observability Model
- FinOps: cost and latency analysis from traces
- AI Agent Center: per-agent view of the same data