Embedding state and detections summary
Not every detection gets an embedding vector: the ingest policy can embed all, some or none at ingest time. An item with no vector cannot be ranked, clustered by similarity or matched by a text search. This page lists where Cropwright shows that and how to repair it.
Where the state appears
- Crop cards and the detail panel show a badge when an item has no vector, and why: encoder failed, deferred, or not selected by the policy. Items with a vector, and items written before the field existed, show no badge.
- Dashboard has an Embedding card in the dataset stats (embedded and not embedded, with the per-state breakdown), and a Detections summary panel below it: totals, an embedding breakdown and a per-label table.
- Projects has an "Embedded" column on each row.
- Ingest results list, per file and in total, how many detections were embedded, not embedded, failed to embed or were filtered out.
Filling in missing vectors
Wherever missing vectors matter there is an Embed action, and each one runs a Reprocess dialog: a dry run first, then the real run behind a confirmation.
- The dashboard's Detections summary offers Embed N detections when the backend suggests it.
- Semantic-search results and ordered views of a cluster show "N items in scope have no vector and are not ranked" with the same action.
- An empty review queue whose reason mentions vectors offers it too.
- Reprocessing a chosen batch of crops with the
embedscope offers "Only items without a vector" and which parts to embed. - The auto-label run on the dashboard has an Embed missing vectors first checkbox, and its progress shows an "embedding items without a vector" stage.
Region writes
After you edit or accept region boxes, the backend reports how many boxes
still lack a vector. /review shows "N boxes have no vector yet" with an
Embed now action while that number is above zero.