The Overview dashboard
Before you start
- A project that has received traces (see the Quickstart)
- Any role in the project can open the Overview
The Overview is the first page of a project. It answers one question for a window of time: how are my models doing? It shows what you spent, how fast and how reliable the calls were, and the newest traces and errors. Every card follows the same time range and environment.
Choose the window
Section titled “Choose the window”Use the controls at the top right of the page, or in a row above the page on a phone.
- Time range. Pick 1h, 24h, 7d or 30d, or use the calendar button for a custom range of up to 90 days. The default is the last 24 hours.
- Environment. Pick one environment or All environments.
Both are kept in the page address and shared with Traces, Sessions and Releases. The line above the greeting names the date, the environment and the range.
When the window includes a whole hour that the worker has already rolled up, as the default 24 hours does, percentiles come from hourly latency histograms and are estimates. An Estimated · hourly rollups chip appears on that line, and the percentile values start with ≈. The tooltip reads “Percentiles are estimated from hourly latency histograms, accurate to about ±25 %. Counts and cost are exact; a span that arrives late can take up to an hour to appear.” A window with no whole rolled-up hour, such as 1h, reads only the raw spans, so its percentiles are exact and the chip does not appear.
The headline
Section titled “The headline”Under the date is a greeting and one sentence about the window.
| Sentence | When |
|---|---|
| “Your models are behaving.” | There are traces and no failed LLM calls. |
| “3 errors need a look.” | At least one LLM call failed. The number is the count of failed LLM calls. |
| “No traces in this window.” | The project has traces, but none in this window. |
| “Waiting for your first trace.” | The project has never received a trace. |
Explore traces opens the Traces page on the same window and environment.
The dark Spend card is the page’s headline.
- The amount is the cost of all spans in the window, with the cents muted. The pill beside the title is the change against the previous window of the same length. A rise in spend is shown as bad news.
- The projection, such as “Projected $5.4K this month”, is the window’s spend per day multiplied by the number of days in the current month. It appears only for windows of 7 days or less, and it is an estimate, not a forecast. The line also shows what the previous window cost.
- The unpriced note, such as “12 calls unpriced · spend is a lower bound”, appears when LLM calls had no price. Those calls add nothing to spend.
- The bars show spend per hour for windows up to two days, and per day for longer ones. The highest bar is lime and carries its value. For the preset ranges the axis counts back from now, for example “−6h” and “Now”. A custom range shows dates instead. Hover a bar for the exact range and amount.
- The three tiles under the bars are the most expensive models, with their dollars and share of priced spend.
If no model in the window has a price, the amount is — with a tooltip, never $0. See how cost is calculated.
Key metrics
Section titled “Key metrics”Four cards sit beside the chart. Each shows a value, a detail line, the change against the previous window of the same length, and a sparkline of the last eight buckets.
| Card | Value | Detail line | Change shown as |
|---|---|---|---|
| p95 latency | 95th percentile duration of LLM calls | The p50 | Milliseconds. A rise is bad. |
| Error rate | Failed LLM calls divided by LLM calls | “203 of 48.2K calls” | Percentage points. A rise is bad. |
| LLM calls | Number of LLM calls | “across 12K traces” | Percent. Neutral. |
| Tokens | Input plus output tokens | Input and output counts | Percent. Neutral. |
These are the definitions behind them:
- LLM calls, errors, error rate, p50 and p95 are computed over spans of kind
llmthat started in the window. Error rate and percentiles are unknown, and shown as —, when the window has no LLM calls. - Cost and tokens are summed over all spans in the window, so they match the trace totals.
- Traces counts the traces that started in the window.
- The environment filter applies to the trace’s environment.
- The previous window is the one of equal length that ends where this one starts.
Calls and errors
Section titled “Calls and errors”The chart shows LLM calls per hour or day as bars, with the failed calls stacked on top of the ones that succeeded. Badges above the chart give the totals. A bucket the window cuts off at the start, or one that is still filling, is drawn lighter and labelled Partial hour (or Partial day) in the legend. Hover for a bucket’s calls, errors and error percentage.
Provider incidents in the window are shaded behind the bars. Show incidents turns the shading off or on, and the choice is remembered in your browser. Show table lists the same numbers, and the incidents, per bucket.
Doctor
Section titled “Doctor”The Doctor card shows the Doctor’s score from 0 to 100 as a ring, with a band next to the title: Good at 80 and above, Fair from 50 to 79, Poor below 50. Open the Doctor takes you to the findings.
Next to the ring, one row per part of the score shows its value and how many points it took off, out of the most it can take: Findings (open findings, up to 40), Errors (error rate, up to 30), Latency (p95 against the baseline, up to 15) and Cost (spend against the baseline, up to 15). The baseline is the window of equal length before this one. The score is 100 minus those penalties; hover or focus the ring for the same breakdown.
Under the rows, the card shows the most recent open finding, when there is one, then when the detectors last ran (“Last check 9 minutes ago · 15 detectors ran · 0 failed”) and the status of the providers your calls went to.
If the window has no LLM calls, the score is — with the tooltip “No LLM calls in this window”, and the band reads No score. Open findings are still counted.
Latest traces and recent errors
Section titled “Latest traces and recent errors”Latest traces lists the five newest traces in the window, with name, model, duration, tokens, cost and age. A failed trace has a red tint. For the preset ranges the list shows Live and refreshes every 15 seconds. A custom range is in the past, so it is fetched once. View all opens the Traces page.
Recent errors lists the three newest traces with a failed span, each with its first error message, model and number of failed spans. A badge such as “3+” means more exist. View all opens Traces with the Status: Error filter set.
The two cards count differently from the headline. The headline counts failed LLM calls. Recent errors lists traces where any span failed, such as a tool or an HTTP call.
Models
Section titled “Models”The Models card lists each model with its provider, calls, errors, p50 and p95 latency, input and output tokens, and cost, busiest first. The five busiest show at first. Choose All models to see the rest. Click a model name to open the Traces page filtered to that model. On a narrow card each model becomes a tile with calls, p95 and errors.
A cost of — means the model has no price. Its calls are not part of spend.
Check it worked
Section titled “Check it worked”Send a trace to the project and open Overview. The page leaves the first-run screen and shows the dashboard. The headline reads “Your models are behaving.” when the window has traces and no failed LLM calls, or “N errors need a look.” (with one failure, “1 error needs a look.”) when some failed, and the Latest traces card lists your trace.
Troubleshooting
Section titled “Troubleshooting”- The page shows “Connect your app” with a “Listening for new traces” pill. The project has never received a trace. Send one, for example with the Quickstart. The page checks every three seconds and switches to the dashboard when the first trace arrives. If none arrives, see No trace arrives. The full setup flow behind Connect your app adds two more snippet tabs, curl and No SDK (the gateway). Outside the demo workspace, Try the live demo opens a shared, read-only sample workspace and signs you out of your own, so it asks first.
- The page says “No traces in this time range”. The project has traces, but none in this window or environment. Select Show last 30 days to widen the range, or All environments to clear the environment. The first button is hidden when the range is already 30 days.
- One card shows an error message and a retry button. That card’s request failed. Each card has its own message, so one failed request does not blank the page. Select the retry button.
- A cost shows “—”. The model has no known price, so its calls are not part of spend. See how cost is calculated.