fixes

app.py

@@ -25,20 +25,20 @@ def respond(
         "top_p": top_p
     }
 )
-    try:
-        output = response.json()["choices"][0]["text"]
-    except requests.exceptions.JSONDecodeError as e:
-        cleaned_output = response.text.strip().rstrip('%')
-        output = json.loads(cleaned_output)
-
-    if isinstance(output, str):
-        return output
-    elif 'error' in output:
-        if output['error']['code'] == 'concurrency_limit_exceeded':
-            return "Error: Concurrency limit exceeded. Try submit your request again in a few seconds."
-        else:
+    if response.status_code == 200:
+        return response.json()["choices"][0]["text"]
+    elif response.status_code == 429:
+        try:
+            output = response.json()
+            return f"Error: {output['error']['code']}. Try submit your request again in a few seconds."
+        except requests.exceptions.JSONDecodeError as e:
+            return f"Error: {response}. Refresh the page and try again later, post the error on OLAT forum if your issue persists."
+    else:
+        try:
+            output = response.json()
             return f"Error: {output['error']['message']}. Refresh the page and try again later, post the error on OLAT forum if your issue persists."
-
+        except requests.exceptions.JSONDecodeError as e:
+            return f"Error: {response}. Refresh the page and try again later, post the error on OLAT forum if your issue persists."
 
 demo = gr.Interface(
     fn=respond,