Float 16, and better polling

app.py

@@ -587,14 +587,24 @@ def create_app():
                 if (window.jobPollInterval) {
                     clearInterval(window.jobPollInterval);
                 }
+                console.log('[Auto-poll] Starting polling every 5 seconds');
                 // Auto-click the check status button every 5 seconds
                 window.jobPollInterval = setInterval(() => {
-                    const checkBtn = document.querySelector('button:has(span:contains("Check Status"))') ||
-                                     Array.from(document.querySelectorAll('button')).find(b => b.textContent.includes('Check Status'));
+                    // Find button by looking for "Check Status" text
+                    const buttons = document.querySelectorAll('button');
+                    let checkBtn = null;
+                    for (const btn of buttons) {
+                        if (btn.textContent.includes('Check Status')) {
+                            checkBtn = btn;
+                            break;
+                        }
+                    }
                     if (checkBtn && checkBtn.offsetParent !== null) {
+                        console.log('[Auto-poll] Clicking Check Status button');
                         checkBtn.click();
                     } else {
                         // Button is hidden (job done), stop polling
+                        console.log('[Auto-poll] Button not visible, stopping');
                         clearInterval(window.jobPollInterval);
                         window.jobPollInterval = null;
                     }

models.py

@@ -248,9 +248,17 @@ class EmbeddingModel(BaseEmbeddingModel):
             "passage_prefix": "",
         })
         
-        # Load the model
+        # Load the model with float16 on CUDA to save VRAM
+        # (12B model: float32 = 48GB, float16 = 24GB)
         print(f"Loading model: {model_id} on {device}")
-        self.model = SentenceTransformer(model_id, device=device)
+        if device == "cuda":
+            self.model = SentenceTransformer(
+                model_id,
+                device=device,
+                model_kwargs={"torch_dtype": torch.float16},
+            )
+        else:
+            self.model = SentenceTransformer(model_id, device=device)
         
         # Set max sequence length if supported
         if hasattr(self.model, "max_seq_length"):