→ חזרה

יומן שינויים

v1.2.0

  • Lost-dog page is now an owner hub: status pill (Looking / Found / Archived), share button, and inline edit.
  • “Further candidates nearby” — a second list of lower-confidence sightings within ~10 km of the last-seen point. Lets owners catch close-by maybes without flooding the primary list with false alarms.
  • Reject sightings (“Not my dog”) so the same suggestion doesn’t reappear.
  • Add another reference photo to an existing dog — improves future matching.
  • Mark a dog as found (with a confirm dialog). Found / archived dogs drop out of spotter results, but their page stays viewable so shared posters keep working.
  • “What’s new since your last visit” banner counting fresh spottings.
  • Backend: dogs gain a status state machine and an inactivity timestamp; archived dogs are excluded from /search results and the recent-list.

v1.1.0

  • Live match-strength meter on the lost-dog form, showing real top-1 accuracy from our eval (~73% with 1 photo, ~86% with 2, best at 3+).
  • Match cards now show eval-calibrated Strong / Possible / Weak confidence pills, derived from the open-set ROC of the production matcher.
  • Possible-spottings cards show the confidence label alongside the % score so 65% and 85% are visually distinguishable.
  • Internal: offline eval harness (closed-set R@1/R@5/mAP + open-set sensitivity/specificity) shipped as scripts/eval.py and scripts/plot_eval.py — see docs/eval.md.

v1.0.0

  • First public release.
  • Hebrew + English UI with RTL support; Hebrew is the default.
  • Light + dark themes (system preference by default; manual toggle).
  • Footer with version, changelog link, and disclaimer.

v0.5.0

  • Lost-dog cards now surface possible spottings from user reports.
  • Spotter results include matches from other user sightings.
  • Owners can confirm sightings to grow a dog's reference cluster.

v0.4.0

  • Pivoted to cluster model: lost dogs are 1–5 photos as one identity.
  • Multi-query max-similarity ranking with cluster bonus for corroborated matches.
  • Duplicate detection on registration (similarity ≥ 0.95).
  • Contact split into name (required), email (required), phone (optional).
  • Recently-lost strip on the home page.

v0.3.0

  • Cloud deployment: Vercel + Hugging Face Spaces + Supabase.
  • Dual-backend storage (MinIO local, Supabase Storage in production).
  • Snapshot exporter/importer for portable seed data.

v0.2.0

  • Multiple images per search using Option B max-similarity.
  • Last-seen-from-spotter distance per match.
  • Image format + size validation (JPEG/PNG/WebP, ≤10 MB).
  • Camera / gallery / files action sheet for mobile.

v0.1.0

  • First end-to-end working build.
  • YOLOv8 detection + DINOv2 embedding + pgvector search.
  • Tel Aviv as default geographic center.