fixed bug for Mistral RAG/chat template

app.py

@@ -86,7 +86,14 @@ else:
 model = model.to(device)
 
 
-def chat(message, history, temperature, repetition_penalty, do_sample, max_tokens):
+def chat(
+    message,
+    history,
+    temperature=0,
+    repetition_penalty=1.1,
+    do_sample=True,
+    max_tokens=1024,
+):
     print("repetition_penalty:", repetition_penalty)
     chat = []
     for item in history:
@@ -136,9 +143,12 @@ def chat(message, history, temperature, repetition_penalty, do_sample, max_token
     partial_text += f"1. Total Repetitions: {total_repetitions:.3f}\n"
 
     if index >= 0:  # RAG
-        scores = calc_bleu_rouge_scores(
-            [answer], [questions[index]["wellFormedAnswers"]], debug=True
+        key = (
+            "wellFormedAnswers"
+            if "wellFormedAnswers" in questions[index]
+            else "answers"
         )
+        scores = calc_bleu_rouge_scores([answer], [questions[index][key]], debug=True)
 
         partial_text += "\n\n Performance Metrics:\n"
         partial_text += f'1. BLEU: {scores["bleu_scores"]["bleu"]:.3f}\n'
@@ -150,6 +160,7 @@ def chat(message, history, temperature, repetition_penalty, do_sample, max_token
 demo = gr.ChatInterface(
     fn=chat,
     examples=examples,
+    cache_examples=False,
     additional_inputs_accordion=gr.Accordion(
         label="⚙️ Parameters", open=False, render=False
     ),

app_modules/llm_inference.py

@@ -166,7 +166,7 @@ class LLMInference(metaclass=abc.ABCMeta):
     def apply_chat_template(self, user_message):
         result = (
             []
-            if self.llm_loader.model_name.lower().startswith("gemma")
+            if re.search(r"gemma|mistral", self.llm_loader.model_name, re.IGNORECASE)
             else [
                 {
                     "role": "system",