Update main.py

main.py

@@ -209,30 +209,26 @@ async def search_database(file: UploadFile = File(...), detect_faces: bool = For
                 is_face = vec_dict["type"] == "face"
 
                 # ── Score filtering ──────────────────────────────────────
-                # Face lane: RetinaFace+GhostFaceNet 512-D cosine similarity.
-                # Raw scores cluster around 0.3-0.5 for same person; anything
-                # below 0.30 is noise. We remap to 75-99% for the UI.
+                # Face lane: GhostFaceNet 512-D cosine similarity.
+                # Raw scores 0.3-0.5 = same person. Remap to 75-99% for UI.
                 if is_face:
-                    RAW_THRESHOLD = 0.30
+                    RAW_THRESHOLD = 0.35   # matches original cloud_db.py
                     if score < RAW_THRESHOLD:
                         continue
                     ui_score = 0.75 + ((score - RAW_THRESHOLD) / (1.0 - RAW_THRESHOLD)) * 0.24
                     ui_score = min(0.99, ui_score)
                 else:
                     # Object lane: SigLIP+DINOv2 1536-D fused cosine similarity.
-                    # These vectors are dense and high-dimensional — even completely
-                    # different images can score 0.70-0.80 just by chance.
-                    # 0.82+ = genuinely similar images (same animal, same scene type)
-                    # 0.70-0.82 = loosely related (outdoor vs outdoor, animal vs animal)
-                    # < 0.70 = unrelated — hard filter these out entirely
-                    MIN_OBJECT_SCORE = 0.70
+                    # Matches original cloud_db.py min_score=0.45 floor.
+                    # Scores below 0.45 are pure noise — unrelated images.
+                    MIN_OBJECT_SCORE = 0.45
                     if score < MIN_OBJECT_SCORE:
                         continue
                     ui_score = score
 
                 caption = "👤 Verified Identity" if is_face else match["metadata"].get("folder", "🎯 Object Match")
                 out.append({
-                    "url":     match["metadata"].get("url", ""),
+                    "url":     match["metadata"].get("url") or match["metadata"].get("image_url", ""),  # "image_url" = legacy key from cloud_db.py
                     "score":   round(ui_score, 4),
                     "caption": caption,
                 })