Fixed a critical bug preventing the offline AI model from loading in production. Added proper access rights, so the text predictor now works flawlessly on all websites.