Update app.py

app.py

@@ -31,21 +31,20 @@ async def chat(request: Request):
 
         output_stream = llm(prompt, max_tokens=1024, echo=False, temperature=0.2, top_p=0.1, stream=True)
 
-        response_text_chunk = ''
-
+        # Remove the response_text variable and directly print the chunks
         while True:
             try:
                 chunk = next(output_stream)
                 if chunk.get('choices') and chunk['choices'][0].get('text'):
-                    response_text_chunk += chunk['choices'][0]['text']
-                    typewrite(response_text_chunk, delay=0.00)
+                    text = chunk['choices'][0]['text']
+                    typewrite(text, delay=0.00)
             except StopIteration:
                 break
 
         if ask == 'clear':
             os.system("cls")
 
-        return {"response": response_text_chunk}
+        return {} # Since response_text is no longer used, return an empty dictionary
 
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))