Update app.py

app.py

@@ -3636,12 +3636,26 @@ def respond(message, history, max_tokens, temperature, top_p):
             messages.append(["assistant", val["content"]])
     messages.append(["user", message])
     
-    response = ""
-    for message in client.chat_completion(
-        messages, max_tokens=max_tokens, stream=True, temperature=temperature, top_p=top_p
-    ):
-        token = message.choices[0].delta.content
-        response += token
+    max_retries = 25 
+    retry_count = 0
+
+    while retry_count < max_retries:
+        try:
+            response = ""
+            for message in client.chat_completion(
+                messages, max_tokens=max_tokens, stream=True, temperature=temperature, top_p=top_p
+            ):
+                token = message.choices[0].delta.content
+                response += token
+                yield response  
+            break  
+        except Exception as e:
+            print(f"Attempt {retry_count + 1} failed: {e}")
+            retry_count += 1
+            time.sleep(1)  
+
+    if retry_count == max_retries:
+        response = random.choice(errored_responses)
         yield response