testing padding impact (just for debug - will be removed later)

app.py

@@ -3,7 +3,7 @@ import datasets
 import faiss
 import os
 
-from transformers import pipeline  # , AutoModel, AutoTokenizer
+from transformers import pipeline
 
 
 auth_token = os.environ.get("CLARIN_KNEXT")
@@ -34,20 +34,16 @@ def load_index(index_data: str = "clarin-knext/entity-linking-index"):
 
 
 def load_model(model_name: str = "clarin-knext/entity-linking-encoder"):
-    # tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=auth_token)
-    # model = AutoModel.from_pretrained(model_name, use_auth_token=auth_token)
-    model = pipeline("feature-extraction", model="clarin-knext/entity-linking-encoder", use_auth_token=auth_token)
-    # return tokenizer, model
+    model = pipeline("feature-extraction", model=model_name, use_auth_token=auth_token)
     return model
 
 
-# tokenizer, model = load_model()
 model = load_model()
 index = load_index()
 
 
 def predict(text: str = sample_text, top_k: int=3):
-    # query = prepare_query(tokenizer, text)
+    text = text + "".join(['[PAD]' * 252])
     index_data, faiss_index = index
     # takes only the [CLS] embedding (for now)
     query = model(text, return_tensors='pt')[0][0].numpy().reshape(1, -1)