Analytics

Watch the workflow: Follow an analytics category to recorded evidence · Authored dummy data

Switch an authored source-defined event analysis to category totals, filter a six-event category, inspect one exact record reference and follow its source trace. These are retained event counts, not customer outcomes or maintained-catalog approval.

Read the walkthrough transcript

• 00:00 — Compare authored event counts in Tensile. This demo uses a source-defined analysis.
• 00:06 — Select high to inspect six authored events. Counts describe retained evidence only.
• 00:11 — Open an exact record reference. See the recorded source trace behind the count.
• 00:17 — Follow the source trace. Recorded events do not establish customer success.

Open Observability in the sidebar, then select Analytics. For a connected source with maintained dimensions, choose from the approved fields already available for that workspace. A saved analysis loads automatically. Analytics uses its own selected dates; global data-view filters do not apply.

Explore findingsCopy Anchor Link

Dates offers the last 24 hours, 7 days, 30 days, or a custom UTC range of up to 31 days. Presets outside retained history are disabled. Custom dates explain the earliest available date; a rejected older window offers Choose dates instead of repeating the same read. Group by selects an approved dimension; Split by optionally adds a second. Changes update the result automatically. View switches between a time trend, category totals and a two-dimensional cross-tab when a split is selected. Settings contains saved analyses and evidence coverage.

Maintained dimensions come from customer-specific metadata. Safe fields with two to eleven values can be approved for grouping; identifiers, person names, contact details, free text and measurements are excluded. New codes remain pending until reviewed. Ordinary readers see pending counts without the raw codes. See Maintained Analytics dimensions for how those fields are reviewed and updated.

From a trend to a traceCopy Anchor Link

Each line represents a category. A finding is counted once, at its first eligible event. Repeated rows are deduplicated. Counts describe observed stored evidence; they do not establish complete source coverage or correctness.

Select a legend category to filter its records across the window. Select a time point to filter to that category and time bucket. For dense or overlapping points, use Inspect time to choose a time bucket and then its category. The category buttons show exact counts, including zero. Selected filters appear above the records table. Clear filters restores the full observed cohort and moves keyboard focus back to the records.

Points support Enter and Space to select evidence, arrow keys to move between dates and categories, and Home and End for the first and last date. The readout shows the exact category, date and count. Partial edge buckets are labelled.

The records table shows time, selected grouping fields, record reference and a direct View trace link. Choose a record reference to inspect its source traces. Source links retain the exact native customer and session. Returning to a saved analysis restores its dates and selected evidence. An unavailable or changed reference is reported explicitly; another customer's reused session is never substituted.

Save and recoverCopy Anchor Link

Use Settings → Saved analyses to select another analysis or change its name. Save analysis stores a personal definition in your current workspace; existing saved entries remain intact. New analysis stays with the connected source's maintained dimensions. If a saved version is unavailable, New analysis with current fields keeps the known analysis name, source and dates in a separate current draft; it does not reactivate the unavailable saved version. Saving requires member, admin or owner permission. Unsaved drafts use workspace/user-scoped browser session storage; source results are not stored there.

Personal maintained analyses retain their approved mapping. They cannot silently adopt later category meanings. Withdrawn or unsafe fields make an affected analysis unavailable. Choose another approved analysis or start a new one. If no approved fields remain, values need review before the analysis can run. An unavailable catalog reports a held or unavailable state instead of opening a manual field editor or showing partial totals.

Braintrust analytics includes all business accounts in the configured project connected to your workspace; it does not require a manual per-account grant. Account identities still keep records separate. Reads and saves recheck workspace/source access and held-out evidence. For a transient read failure, use Retry; for a changed mapping, use Load current mapping. For an available analysis, Show all records clears an invalid incoming selection while preserving the dates. An unavailable saved version instead offers another available analysis or a new analysis with current fields.

Existing source-defined analysesCopy Anchor Link

Older saved analyses can use a source-specific definition rather than maintained dimensions. Their Settings additionally exposes field discovery, count unit, stable keys, cohort event, enum codes and a recorded-response mapping. Confirm those assumptions before applying an edited definition. This is a supported existing mode; it is not a recovery path for an unavailable maintained catalog.

In this mode, discovery shows field coverage and counts rather than arbitrary raw values. A newly discovered field requires confirmed non-sensitive enum codes before use. Include later responses can extend response matching up to 90 days beyond the selected end time. An answered finding has a matching recorded response; missing responses remain Unknown when the source cannot certify complete response coverage. Unknown does not mean Unanswered, and a recorded dismissal or waiver does not prove correctness.

Missing stable identity makes unit totals and trends unavailable. Evidence & coverage explains identity and response gaps; Excluded events provides exact evidence links for excluded rows. Oversized, failed or inconsistent reads report an error instead of partial metrics. Use a smaller range when needed.

Analytics makes no model or provider calls. Ingestion, corrections and retention can change observed counts; selected dates fix event-time cutoffs rather than an immutable ingestion snapshot.