Add name-based deduplication to prevent duplicate label names

handler.py

@@ -152,12 +152,32 @@ class EndpointHandler:
         if not new_items:
             return 0
         with self._lock:
-            known = set(getattr(self, "class_ids", []))
-            batch = [it for it in new_items if int(it.get("id")) not in known]
+            # Get ALL existing IDs and names from current state
+            known_ids = set(getattr(self, "class_ids", []))
+            known_names = set(getattr(self, "class_names", []))
+            
+            # Filter items, checking against both ID and name
+            batch = []
+            for it in new_items:
+                item_id = int(it.get("id"))
+                item_name = it.get("name")
+                
+                # Skip if either ID or name already exists
+                if item_id in known_ids:
+                    continue  # Skip duplicate ID
+                elif item_name in known_names:
+                    continue  # Skip duplicate name
+                else:
+                    batch.append(it)
+            
             if not batch:
                 return 0
+                
+            # Process the filtered batch
             prompts = [it["prompt"] for it in batch]
             feats = self._encode_text(prompts).detach().cpu().to(torch.float32)
+            
+            # Update the persistent state
             if not hasattr(self, "text_features_cpu"):
                 self.text_features_cpu = feats.contiguous()
                 self.class_ids = [int(it["id"]) for it in batch]
@@ -166,6 +186,7 @@ class EndpointHandler:
                 self.text_features_cpu = torch.cat([self.text_features_cpu, feats], dim=0).contiguous()
                 self.class_ids.extend([int(it["id"]) for it in batch])
                 self.class_names.extend([it["name"] for it in batch])
+            
             self._to_device()
             return len(batch)