Update rag.py

rag.py

@@ -54,6 +54,7 @@ class RAG:
             max_tokens=model_parameters['max_new_tokens'], # TODO: map other parameters
             frequency_penalty=model_parameters['repetition_penalty'], # this doesn't appear to do much, not a replacement for repetition penalty
             #repetition_penalty=model_parameters['repetition_penalty'], # TODO: figure out how to pass repetition penalty
+            extra_body={'repetition_penalty': model_parameters['repetition_penalty']},
             stream=False
         )
         print(chat_completion)