Update app.py

app.py

@@ -13,44 +13,34 @@ llm = Llama(
 def format_chat_history(history):
     formatted_history = ""
     for user_msg, assistant_msg in history:
-        formatted_history += f"Pitanje: {user_msg}\nOdgovor: {assistant_msg}\n"
+        formatted_history += f"Q: {user_msg}\nA: {assistant_msg}\n"
     return formatted_history
 
-def clean_text(text):
-    # Basic cleaning
-    text = text.replace("ASSISTANT:", "").strip()
+def process_chunk(text):
+    # Remove any special tokens
+    text = text.replace("ASSISTANT:", "").replace("A:", "")
     
-    # Handle spaces after punctuation
-    punctuation_marks = ['.', ',', '!', '?', ':', ';']
-    for mark in punctuation_marks:
-        text = text.replace(mark, mark + ' ')
+    # Add space after each character if there isn't one already
+    processed_text = ""
+    for i, char in enumerate(text):
+        processed_text += char
+        if i < len(text) - 1 and not text[i+1].isspace():
+            processed_text += " "
     
-    # Normalize spaces
-    words = text.split()
-    text = ' '.join(words)
-    
-    return text
+    return processed_text.strip()
 
 def chat(message, history):
-    system_prompt = """Ti si YugoGPT, visoko precizan AI asistent.
-
-OSNOVNI PRINCIPI:
-- Dajem konkretne i tačne informacije
-- Odgovaram samo o temama koje dobro poznajem
-- Koristim jasan i precizan srpski jezik
-- Fokusiram se na činjenice
-- Odgovaram direktno i pozitivno
-- Izbegavam nagađanja"""
+    system_prompt = """Ti si YugoGPT.
+Odgovaraj direktno i precizno na srpskom jeziku."""
     
     chat_history = format_chat_history(history)
     
     full_prompt = f"""SYSTEM: {system_prompt}
 
-KONTEKST:
 {chat_history}
 
-Pitanje: {message}
-Odgovor:"""
+Q: {message}
+A:"""
     
     response = llm(
         full_prompt,
@@ -59,25 +49,25 @@ Odgovor:"""
         top_p=0.1,
         repeat_penalty=1.2,
         top_k=20,
-        stop=["Pitanje:", "\n\n"],
+        stop=["Q:", "\n\n"],
         stream=True
     )
     
     partial_message = ""
     for chunk in response:
         if chunk and chunk['choices'][0]['text']:
-            text = clean_text(chunk['choices'][0]['text'])
-            partial_message = clean_text(partial_message + text)
+            text = process_chunk(chunk['choices'][0]['text'])
+            partial_message += text
             yield partial_message
 
 demo = gr.ChatInterface(
     fn=chat,
-    title="YugoGPT Stručni Asistent",
-    description="Precizan izvor informacija. PAŽNJA!!! ZNA DA LAŽE!!!",
+    title="YugoGPT",
+    description="Stručni asistent",
     examples=[
-        "Objasnite princip rada relacione baze podataka",
-        "Kako funkcioniše HTTP protokol?",
-        "Opišite osnovne komponente računara"
+        "Objasni kako radi HTTP protokol",
+        "Šta je SQL?",
+        "Kako radi TCP/IP?"
     ]
 )
 
@@ -88,3 +78,4 @@ if __name__ == "__main__":
         share=False
     )
 
+