Use the Console
Find your way around Console
Use Overview to get started, Traces to inspect requests, and Lab to compare and review changes. Project selection determines which evidence you see.
Updated
Applies to Console v0.2.2 with the current navigation. Available actions depend on project access and runtime configuration.
Start with the right organisation and project
Open Console and check both selectors in the top bar. Keep the same project selected while you move between keys, traces, datasets, and evaluations. A page that looks empty may simply belong to a different project.
In shared Console links, keep the existing ?project= value. To find your project ID, open Settings → Project → General. The top search can find traces, agents, keys, projects, prompts, datasets, and evaluations.
Where each task lives
| You want to… | Open | What to check |
|---|---|---|
| Connect an application | Overview → Send a request | The selected project, key, model, and client example. |
| Inspect an individual request | Traces | Status, captured input/output, route, usage, and timing. |
| Compare a prompt or model | Lab → New evaluation | The same cases and evaluator for each candidate. |
| Get a proposed prompt improvement | Lab → New prompt evaluation | Usable traces and registered production prompt lineage. |
| Inspect prompt versions | Lab → Prompts | Live label, candidate version, and version history. |
| Collect evaluation cases | Lab → Datasets | Captured content, cleaning state, and cases actually in Lab. |
| Review outputs and proposed changes | Lab → Review | Suggestions and Human review are separate sections. |
| Inspect deployed agents | Agents | The intended agent and its current state. |
| Choose or test a model | Models → model row | Enabled state, modality, compute readiness, and run cost. |
| Manage credentials and routing | Settings → Project → API keys / Routing | Project scope and enabled routes. |
| Set retention and privacy behavior | Settings → Project → Data policy | Identifier removal and the applicable organisation masking policy. |
| Control spending | Settings → Project → Budget; Organisation → Billing and Pro tokens | Project limits and organisation-level balance/charges. |
| Pair or inspect compute | Settings → Organisation → Compute | Reachability, online status, capabilities, and model assignment. |
Older documentation may group features as Engineering, Lab, and Compute. In the current layout, Engineering is not a sidebar item, and Compute is inside Settings. Navigation layout and release version are separate: some accounts may still use the older layout.
Try a model from its detail panel
- Open Models. Search for the exact model or use a modality filter.
- Click the model row to open its details. Check the provider and whether it needs compute.
- In Test model, enter a small, non-sensitive input or click Load sample. Loading the sample fills the form; it does not run inference.
- Check the input and any media attachment. Available controls differ by modality.
- Click Run model test when ready to make a request. This can consume tokens or provider credit. Read the output or error, then inspect the request in Traces.
The detail panel also shows a Compare with another model control that may be disabled as “Coming soon”. Use Lab → New evaluation → Model for an available comparison workflow.
| State | Meaning | Next step |
|---|---|---|
| Enabled | The model is allowed by the organisation’s model policy. | Check routing and compute before assuming a call will work. |
| Needs compute | The route needs a compatible runtime. | Follow Pair local compute or ask an operator for available managed compute. |
| Starting worker | Setup has begun but the worker is not ready. | Wait for its actual status; inspect failures before retrying. |
| Worker failed | Worker setup did not complete. | Inspect the reported error and runtime requirements. |
| Ready | Compute reports readiness. | Verify a request and runtime evidence for the intended model. |
Read a trace without changing it
Use All / Failed, Tag, Model, Key, Time window, and search to narrow the list. Open a row to inspect its details and events. The Live switch controls live updates.
More actions → Send again makes a new request; it is not a refresh. Flag for review queues human work. Leave those actions alone when you only want to inspect a trace.
If the body says payloads were not captured, the row only proves metered activity. Do not copy that placeholder into a candidate prompt or treat its presence in a matching count as evidence that it is usable for evaluation.
When a page or action is blocked
| Symptom | Check first |
|---|---|
| No traces or datasets | Organisation, project, tag, and dates. Broaden one filter at a time. |
| Unexpected keyword argument tags | Use StimulirClient(tags=[...]) or agent(tag="..."). The current Send a request Python snippet may still show the unsupported per-call tags argument. |
| A model is visible but requests fail | Enablement, project routing, provider credentials, budget, and compute readiness. |
| Evaluation count looks right but input is a placeholder | Open the trace. Capture real input/output and rebuild the cohort before running. |
| Review is empty | Check Pending / Approved / Corrected / Rejected and the project’s Review rules. Review is a queue, not a copy of all traces. |
Continue with your first request, Evaluate and review, or Pair local compute.