A year of Hybrid ML/AI cataloging aid in Archipelago Commons: The state, the lessons and probable future(s) explained through a real production implementation

During OR2024 we showcased our first release of a public ML/AI Image Similarity and semantic search prototype integrated into Archipelago Commons(1.4.0). A primer in OSS Repository systems. We also gave an introduction to the challenges, concerns and hows of our development, including a not so brief 101 on ML inference. A year of research later, we want to share the evolved state of the solution and what will be available for 1.5.0, all this by using a specific community prototype as a showcase: Revs Institute. We will explain through this project how we support field specific human knowledge and responsible decision making to enrich unknowns (media/metadata) through transfer learning from well cataloged and known Digital Objects without giving up control to unattended inference. We will also demonstrate how datasets for re-training and refining of ML Models are produced and reinjected through the act of assisted cataloging, allowing a cyclic, self feeding ML system. Finally, and most importantly, we will share a closing and honest discussion of the expectations, limitations, and carbon footprint effects of ML on production: the scope of ML/AI from our and our community's ethical and system's design goals perspective.