Update app.py

app.py

@@ -113,7 +113,7 @@ def generate_embeddings(texts: Union[str, List[str]]) -> Union[List[float], List
                                 embedding = (sum_embeddings / sum_mask).squeeze().cpu().numpy()
                             else:
                                 # Simple mean pooling without attention mask
-                                embedding = outputs.last_hidden_state.mean(dim=1).squeeze().cpu().numpy()
+                        embedding = outputs.last_hidden_state.mean(dim=1).squeeze().cpu().numpy()
                         else:
                             # Fallback to pooled output if available
                             embedding = outputs.pooler_output.squeeze().cpu().numpy()
@@ -123,9 +123,9 @@ def generate_embeddings(texts: Union[str, List[str]]) -> Union[List[float], List
                 elif model and hasattr(model, 'encode'):
                     # Method 2: Using sentence transformer fallback
                     embedding = model.encode(text)
-                    embeddings.append(embedding.tolist())
-                else:
-                    raise Exception("No model available")
+                        embeddings.append(embedding.tolist())
+                    else:
+                        raise Exception("No model available")
                         
             except Exception as e:
                 logger.warning(f"Error generating embedding for text: {str(e)}")
@@ -309,8 +309,12 @@ async def predict(data: dict):
             # Normalize embeddings if requested
             if normalize:
                 import numpy as np
-                embeddings = [emb / np.linalg.norm(emb) for emb in embeddings]
-                logger.info("Embeddings normalized")
+                try:
+                    embeddings = [emb / np.linalg.norm(emb) for emb in embeddings]
+                    logger.info("Embeddings normalized")
+                except Exception as norm_error:
+                    logger.warning(f"Normalization failed: {str(norm_error)}, returning unnormalized embeddings")
+                    # Continue with unnormalized embeddings
             
             return {
                 "embeddings": embeddings,