Optimizing Retrieval-Augmented Generation for Space Mission Design via Multi-Task Learning

The extensive documentation required during the early phases of space mission design, which must be reviewed and analyzed by systems engineers, often slows down the design process. Given the time-intensive nature of this stage, integrating a Virtual Assistant (VA) can significantly enhance the efficiency of information retrieval. Recent advancements in Natural Language Processing (NLP), particularly the development of Large Language Models (LLMs), have created opportunities for non-invasive question-answering VAs. However, the limited technical knowledge of LLMs in the space domain and their lack of up-to-date information pose significant challenges for real-world applications. Retrieval-Augmented Generation (RAG) offers a solution by incorporating an external knowledge source, from which a retriever selects the most relevant information based on the user’s query. In this thesis, we enhance the retriever component by fine-tuning it on space-domain data using supervised and Multi-Task Learning (MTL) approaches, demonstrating how improvements in the retriever impact the performance of the generator. In the MTL setting, we combined supervised with self-supervised tasks showing how the network benefits from the addition of complementary tasks. Finally, we tested the enhanced RAG pipeline during a Concurrent Engineering (CE) session at the Argotec Advanced Concept Laboratory (ACLab), yielding promising results in a real-world scenario.