Update app.py

app.py

@@ -56,7 +56,7 @@ async def predict(request: Request):
         print("OUTPUTS",outputs)
         # Prepare response with converted types
         return {
-            "embedding": outputs[0].astype(float).tolist(),
+            "embedding": outputs[0][0].astype(float).tolist(),
             "input_ids": inputs["input_ids"][0].tolist(),
             "attention_mask": inputs["attention_mask"][0].tolist()
         }