Exploring the interplay between DataSpaces and Large Language Models

The confluence of Large Language Models (LLMs) and Dataspaces presents a captivating prospect for the future of data science and AI, fostering new avenues for data exploration, analysis, and knowledge extraction. This paper explores both patterns arising from the Dataspaces-LLMs convergence: Dataspaces applied to LLMs (DS4LLM) and LLMs applied to Dataspaces (LLM4DS), further investigating the latter. Dataspaces, a conceptual framework for data-centric systems, are poised to revolutionize data management and sharing, indeed. They provide a structured and secure environment for managing and integrating data from multiple sources. Traditional approaches to data management, nevertheless, are bottom-up, often hindering the timely use of data. This paper advocates for a top-down perspective, prioritizing data collection and provisioning to other data management tasks that have to be thus delivered on purpose and in a timely manner a-posteriori, after data (re)source discovery. To bridge the gap between these two approaches, LLMs emerge as a powerful tool able to unlock the value hidden within vast data repositories and Dataspaces. By LLMs, data management tasks such as filtering, cleaning, aggregation, integration, and augmentation can be automated into LLM4DS workflows. This research specifies some LLM4DS prompt templates that can be tailored to specific use cases for on demand and on purpose (a-posteriori) data management requiring post-scheme. A case study on a medical Dataspace demonstrates the practical application and suitability of the proposed LLM4DS approach to the problem at hand.