Update app.py

app.py

@@ -33,19 +33,22 @@ async def chat(request: Request):
 
         response_text_chunk = ''
 
-        while True:
+        async def generate_chunks():
+            nonlocal response_text_chunk
             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)
+                while True:
+                    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)
+                        yield {"response": response_text_chunk}
             except StopIteration:
-                break
+                pass
 
         if ask == 'clear':
             os.system("cls")
 
-        return {"response": response_text_chunk}
+        return generate_chunks()
 
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))