A 1B model doesn't read your letter: the unglamorous layer between documents and local AI