Dashboards

Overview

Dashboards let you compose widgets — graphs, tables, and big-number tiles — over your project’s traces, threads, metric data, and annotations. Each widget pulls live data, supports filters and breakdowns by dimensions like classifier labels, end users, or metadata keys, and can be exported as CSV, PNG, or PDF for sharing.

A dashboard is a draggable grid of widgets. Resize and rearrange any widget by grabbing its drag handle; layouts are saved per dashboard.

Create a Dashboard

Choose a Widget Shape

When you add a widget, the first thing you pick is the data shape. The shape decides whether your data is bucketed across time or aggregated as a single snapshot.

Once you pick a shape, you’ll see the matching display types in the editor. Switch between them at any time without losing your configuration.

Time-Series Widgets

Time-series widgets answer “how does X change over time?”. They bucket your data into intervals (the bucket size scales with the date range) and plot one value per bucket.

Categorical Widgets

Categorical widgets answer “what’s the distribution / what’s the headline?” over a fixed window — no time buckets.

Configure a Widget

Inside the widget editor, set what data is plotted and how it’s sliced.

Data Model

The Data Model is the source of the values your widget reads. Pick one of:

For Span, Metric Data, and Annotation, the editor reveals nested selectors so you can scope deeper without picking the wrong dimension by accident.

Aggregation

The Aggregation is the value the widget plots — Count, Average latency, P95 latency, Unique end users, Average score, etc. The list of aggregations adapts to the data model you picked, so you only ever see options that make sense.

Filters

Filters narrow the dataset using the same syntax as the Observatory. Match by environment, tag, metadata field, classifier label, score, latency, or any combination — exactly the same expressions you’d build to search for traces.

In Manual mode (see below), each line has its own filters so you can compare different slices side-by-side. In Breakdown mode there’s one shared filter that scopes the whole widget; the editor surfaces a hint to that effect when you toggle modes.

Manual vs Breakdown (time-series widgets)

Time-series widgets have two modes, switchable at the top of the editor:

• Manual lines — define up to 5 lines yourself, each with its own data model, aggregation, and filters. Use this when you want to compare specific things side-by-side — e.g. avg latency for LLM spans vs. avg latency for Retriever spans, or trace count in prod vs. trace count in staging.
• Breakdown — pick one data model + aggregation, then pick a dimension. The widget auto-creates one series per value of that dimension (capped by Top K). Use this when you want to slice one metric by an attribute — e.g. trace count broken down by Failure Mode label or average score broken down by metric collection.

Switching modes resets the configuration so you can start fresh in the new shape.

Breakdown Dimensions

In Breakdown mode (and for categorical widgets), the Dimension controls how the metric is split. The available dimensions adapt to the data model — common ones include:

• End user — split by the endUserId attached to traces or threads.
• Tag — split by trace tag.
• Span type — LLM, Tool, Retriever, Agent, etc.
• Metadata — split by a specific metadata key. The editor prompts you to pick or type the key (see the metadata tip above).
• Classifier — split by the labels of one of your project’s classifiers. Pair this with Signals to graph things like trace count by Failure Mode label over time.

Top K

When a breakdown could produce many series, Top K caps the result:

• Direction — Top (highest by value) or Bottom (lowest).
• Limit — an integer between 1 and the configured ceiling (commonly 20). The remaining values are not plotted; switching the direction or raising the limit always re-evaluates against the underlying data.

Time Range

By default a widget inherits the dashboard’s date range. Toggle Custom time range in the editor to keep that one panel scoped (e.g. always last 7 days, regardless of the dashboard date) — useful for “headline KPI” widgets that should always show a fixed window.

Manage Widgets

Each widget has a kebab menu (⋮) in its top-right corner with the following actions:

PNG and PDF exports use the widget’s current rendered state — title, legend, axes, and data — so the exported file matches what you see. Filenames default to the widget’s name, sanitized for the filesystem.

Manage the Dashboard

The Manage button in the dashboard header opens a dropdown with Edit (rename, change description, flip visibility) and Delete. Deleting a dashboard removes it and all of its widgets — the underlying trace, thread, metric, and annotation data is unaffected.

To rearrange the layout, drag a widget by its drag handle to move it, or drag a widget edge to resize it. Layouts persist as soon as you let go.