Clusters (/p/<project>/clusters, /p/<project>/clusters/[id])
Cluster grid (/p/<project>/clusters)
Every cluster as a card: dominant class, cohesion (the served nearest-centroid purity — labelled "cohesion" rather than "purity" so a mixed-labels-but-visually-coherent cluster doesn't read as "bad"), size and representative crops.
- Filter by cohesion band or class; search by embedding similarity (semantic search, when enabled) or literal OCR text.
- An Ignored toggle lists the excluded bucket with a restore action.
- When the backend serves an embedding projection, an overlay lets you lasso a set of points on a 2-D plot and preview them side by side — visualization only, it never feeds a clustering decision.
- Filtering on the region class swaps the grid for the region gallery (browsing the region queue directly, with detector/verified/status/score/ text filters, plus a Box state filter; see Multi-box regions).
- One synthetic inventory card per registered annotation slot with a browse endpoint is pinned to the unfiltered grid.
The filter bar above the grid, and the Matching items mode that lists and bulk-edits everything the filter matches, are described in Item filter and Matching items.
Cluster detail (/p/<project>/clusters/[id])
The core triage surface: a crop grid for one cluster with pointer-based drag and drop onto class rows, bulk select/confirm/move/discard/ignore, per-class hotkeys, Assist… (ask the VLM about the selection, the cluster or the open sub-cluster, with a per-run endpoint and a proposal for the cluster's label, see Assist (VLM)) and Accept VLM for the page, Reprocess for the selection (see Reprocess), Refine (sub-cluster with agglomerative clustering; a sub-cluster tab lists that sub-cluster's members from the server), a strategy bar (sort, score chips, diverse-selection overlay) and a core-member cut line.
See Keyboard shortcuts for the full key table.
Crop detail
The info button on any crop opens its full provenance: the source image with every item and region box drawn client-side (toggleable), the crop itself, class/label-source/confidence/cluster metadata, OCR text lines, label-write history, and sibling crops from the same source image. It also shows provenance when recorded (import, combine origin, VLM endpoint, model and prompt pack), and a lock icon on a label a human has set.