Expose shared embedder for indexing jobs

scene_search/search_service.py

@@ -38,11 +38,17 @@ class SearchService:
     def model_loaded(self) -> bool:
         return self._embedder is not None
 
-    def _get_embedder(self) -> SiglipEmbedder:
+    def get_embedder(self) -> SiglipEmbedder:
+        """Get the shared embedding model, loading it lazily."""
         if self._embedder is None:
             self._embedder = SiglipEmbedder(self.model_id, self.device)
         return self._embedder
 
+    @property
+    def model_lock(self) -> threading.Lock:
+        """Lock guarding exclusive access to the shared GPU embedding model."""
+        return self._model_lock
+
     def health(self) -> dict[str, Any]:
         init_db(self.db_path)
         with connect(self.db_path) as con:
@@ -174,7 +180,7 @@ class SearchService:
         # SigLIP / PyTorch model execution is kept behind one lock so concurrent
         # API requests do not fight over a single GPU model instance.
         with self._model_lock:
-            embedder = self._get_embedder()
+            embedder = self.get_embedder()
             text_vector = embedder.embed_text(query)
 
             result = self.meili.search_vector(