This paper explores the potential of Decision Transformers (DTs) as a general-purpose framework for addressing combinatorial optimization problems, with a particular focus on the Team Orienteering Problem (TOP) as a case. Drawing on the foundational idea that transformers can model decision-making as a sequence generation task, the TOP is reformulated as a sequential decision process. The work provides a detailed conceptual and methodological foundation for applying transformer-based architectures, which were originally developed for natural language processing, to combinatorial problems. Trained offline on supervised data that includes both typical and high-quality solutions, the DT is able to learn effective construction strategies. The paper concludes with a discussion of important modeling considerations, along with current challenges and potential future directions.