Curation Pipeline

Curation Pipeline A workflow diagram generated by Archify. 01 / Ingest (api) 02 / Curation workers 03 / Review (API + labeling UI) 04 / Export, train, serve Project · /curation/projects/{project} · Ingest (api) Project /curation/projects/{project} Ingest / import · /ingest/*, /datasets/imports · Ingest (api) Ingest / import /ingest/*, /datasets/imports Item detector · OP_INGEST_PRIMARY_* · Ingest (api) Item detector OP_INGEST_PRIMARY_* PE-Core embed · pe_embedding 1024-d · Ingest (api) PE-Core embed pe_embedding 1024-d Config store · profile / pack / VLM · Curation workers Config store profile / pack / VLM Region cascade · per-box verdicts · Curation workers Region cascade per-box verdicts Cluster · IVF + residual AHC · Curation workers Cluster IVF + residual AHC VLM suggest · vlm_* fields · Curation workers VLM suggest vlm_* fields Reprocess · /reprocess (locks kept) · Review (API + labeling UI) Reprocess /reprocess (locks kept) Review queues · /review/{tab} · Review (API + labeling UI) · human label Review queues /review/{tab} human label Export · frozen holdout + YOLO · Export, train, serve Export frozen holdout + YOLO Train · curation-trainer · Export, train, serve Train curation-trainer Bake-off · curation-evaluator · Export, train, serve Bake-off curation-evaluator Promote · Triton load · Export, train, serve Promote Triton load winner active profile item crops job.json bind probe unlocked only validated labels after config change pending_detection Legend HTTP Step Dataset Triton VLM

Bring your own models

  • • You supply the item detector; PE-Core encoders are required
  • • The VLM, region profile and segmenter are optional, and set per project
  • • Ingest 503s until an item detector is configured

Humans stay in charge

  • • VLM answers are suggestions until confirmed
  • • Human and imported labels are locked: no worker or reprocess overwrites them
  • • auto_label clusters only unless run_vlm=true

Closing the loop

  • • Promote is gated on mAP50 and per-class precision/support
  • • Probe predictions feed the model_disagreements queue
  • • Datasets, combine and promoted models match classes by name