This paper presents a multi-level Arabic diacritization system designed to restore diacritics for undiacritized Arabic text. This kind of systems is crucial for Arabic-related NLP tasks and aids learners and individuals with learning difficulties, such as dyslexia or visual impairments. Our system uses a two-level approach: a word-based level and a letter-based level, both employing an encoder-decoder model with a local predictive Luong attention mechanism. The combined model demonstrates good performance with a 22.47% diacritic error rate, significantly surpassing single-level models while maintaining competitive performance despite using a smaller dataset compared to previous studies.