Digital Assistant in the Aviation Industry – Leveraging Generative AI to Enhance Access to Open Data

Sammanfattning

The growth of generative artificial intelligence (GenAI) has rapidly transformed the field of information retrieval over the past few years, particularly through the emergence of large language models (LLMs) and their integration into multimodal systems. This thesis project presents the development of a smart digital assistant that processes flight data from a public API. The aim is to explore how GenAI can be utilized to support operational airport staff by enabling real-time data access and generating responses to user queries related to arrival and departure data. By combining a retrieval-augmented generation (RAG) architecture with dynamic textual and visual outputs, the system delivers real-time information in a format optimized for information retrieval and decision-making within an operational environment. To support visualization generation, two complementary approaches are integrated: automated visualization selection and visualization code generation. These approaches allow the assistant to dynamically render charts in response to natural language (NL) queries. The system is developed and refined through a series of user-centered studies, including a formative study and a task-based evaluation involving operational staff. Results indicate that bar charts and line graphs are the most effective visualization formats for operational tasks.The final evaluation includes a usability study using the System Usability Scale (SUS), the XAI Trust Scale (which assesses trust in explainable artificial intelligence systems), and expert reviews involving both operational staff and stakeholders. Overall, the assistant receives high usability ratings, with stakeholders reporting slightly higher satisfaction and trust levels. Differences in perception reflect distinct focus areas: operational staff emphasize practical effectiveness and interface design, while stakeholders focus on long-term strategic value. Future work should explore combining RAG with fine-tuning to incorporate domain-specific terminology, refining visualization methods, and supporting multi-turn conversations.