Try skipping normalization

Completely off-topic embeddings got a slight relevance increase from
switching to philschmid onnx (0.01 -> 0.03) - curious if this causes it

handler.py

@@ -33,7 +33,5 @@ class EndpointHandler():
         outputs = self.model(**encoded_inputs)
         # Perform pooling
         sentence_embeddings = mean_pooling(outputs, encoded_inputs['attention_mask'])
-        # Normalize embeddings
-        sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
         # postprocess the prediction
         return sentence_embeddings.tolist()