Data sovereignty is the principle that certain information cannot leave the institution that holds it: not to an external server, not to a cloud service. It is not a technicality. An archive with personal files or restricted government records, or contracts with confidentiality clauses, are material that, by data-residency law, by internal policy or by its own content, has to stay inside.
Cataloging with AI runs straight into that. The usual way to use a language model —the kind of artificial intelligence behind tools like ChatGPT— is to send the document to a cloud service, which processes it on a third party’s servers. For restricted material, that is not an option. The question stops being whether AI can catalog and becomes whether it can do so without the documents leaving the institution.
Not every collection has this restriction to the same degree. A library’s catalog is, in the end, public information, so for the general collection sovereignty is rarely the problem. But there is one exception that does bring it to the library: theses and dissertations under embargo. Many, before publication, contain unpublished research, trade secrets or intellectual property not yet protected, and are therefore restricted for a time. Cataloging a thesis with AI requires the model to read the full document, not just the title page, and that document cannot go out to a third-party service, even though the resulting record will be public. It is the case that justifies local processing even where the catalog is open.
Collect was designed for that case. Processing runs on the institution’s infrastructure, on a server or a workstation with a GPU, with a language model that runs locally. Documents are read and turned into records right there; they do not leave. It can run with no internet connection, on an isolated network (what is called air-gapped).
Where an institution wants the quality of a frontier model —the most capable and recent models, which run in the cloud— for the hard cases, there is a hybrid mode. The local model does the first pass, and whatever is escalated to the cloud first goes through an anonymization that replaces the personal data it detects with neutral markers (tokens); on return, they are re-identified. Anonymization reduces the exposure of personal data, but it does not guarantee that none remains, and for those cases the rest of the document’s content does leave the institution. That is why the choice of mode belongs to the institution, according to its policy and its risk: all local, or local with anonymized escalation.
Authority control follows the same principle: it can read authority sources and controlled vocabulary in the cloud, such as id.loc.gov or VIAF, or in the institution’s own systems. With no outside connection, it verifies names and subjects against the authorities and vocabularies the institution has available.
What is rarely found together is neither sovereign AI nor cataloging on their own, but the two at once. Processing on your own infrastructure is offered by several generic providers; producing a catalog record with that is another matter: MARC21, ISAD-G or CDWA, with authority control, deterministic correctors and per-record quality evaluation. Collect does both: the material does not leave, and what it produces is catalogable, not just extracted text.
Two factors weigh on the choice of mode. A local model performs below a frontier one on the hardest cases; the hybrid mode exists for that, and its cost is that an anonymized part of the content leaves the institution. Also, local execution needs a machine with a GPU; it does not run on any computer. With that solved, the material stays inside and the record comes out catalogable.
If your institution has material that cannot leave and needs to describe it, write to us at info@janium.com; we are interested in understanding your restriction, whether data residency, air-gap or confidentiality, and how much room it leaves for quality.