This series crosses two worlds: cataloging —libraries, archives, museums— and artificial intelligence. Each brings its own vocabulary, and not everyone knows both. This is a short reference to the terms that come up most. If you come from a collection, the AI section will help you; if you come from AI, the cataloging one.
Artificial intelligence
Language model (LLM)
The kind of artificial intelligence behind tools like ChatGPT: a program trained on enormous amounts of text that, from what it reads, identifies, summarizes and completes information. “LLM” stands for large language model. In this blog it is what reads the document and proposes the record.
RAG — retrieval-augmented generation
Fragments close to a query are retrieved and handed to a language model so it can draft an answer. It is useful for asking questions in natural language over a set of documents. It does not replace a catalog: it does not give an alphabetical index of authors or subjects, and the amount of text the model can actually use has a ceiling (the effective context window).
Frontier model
The most capable and recent language models, run by large providers and reachable over the internet (in the cloud). They perform better on the hard cases, but processing with them means sending them the material. They contrast with a local model, which runs on the institution’s infrastructure.
Anonymization and tokens
Replacing the personal data in a text with neutral markers —the tokens— before sending it to an external service, and restoring them on return. It reduces the exposure of personal data, though it does not guarantee that none remains.
OCR — optical character recognition
Converting the image of a scanned document into text a machine can read. It is the prior step when the material is a scan with no selectable text.
ASR — automatic speech recognition
Converting speech into text. In audio and video it is what produces the transcript from which the record is extracted. “ASR” stands for automatic speech recognition.
Infrastructure and sovereignty
On-premise
That processing runs on the institution’s own servers, not on a third-party service. The material does not leave.
Air-gapped
An isolated network, with no internet connection. The highest degree of sovereignty: the system works without communicating with the outside.
Data sovereignty and data residency
The principle that certain information cannot leave the institution —or the country— that holds it, by law, by policy or by the type of content. Developed in Cataloging with AI without the material leaving the institution.
Cataloging standards
MARC21 / UNIMARC
The standard format for bibliographic records in libraries. It structures the description into coded fields and subfields.
Dublin Core
A simple set of fields to describe any resource in an interoperable way. Less detailed than MARC, easier to exchange.
ISAD-G
The international standard for archival description. It describes the material in a hierarchy —fonds, section, series, item— where provenance and context matter, not just the isolated piece.
Bubble-up
When describing an archive in ISAD-G, raising to the higher levels of the hierarchy —series, fonds— the data that most of the documents hanging from them share, instead of repeating it in each one. It applies the standard’s non-repetition principle. It is a majority heuristic, with a configurable threshold, and its result is reviewed.
CDWA
The standard for describing works of art and cultural objects, with the Getty vocabulary. It distinguishes the work from its reproduction.
Corporate extensions
Fields Collect adds to the record (on top of Dublin Core or ISAD-G) when the material is a management document: people with their role, organizations, amounts, critical dates, guarantees, clauses, obligations and relationships between entities. In Janium they are indexed as specialized access points. What the document does not state is not filled in.
Access point
A heading by which a record is found in the catalog: a name, a subject, a title, a date. Collect extracts them from the material (it surfaces them). In Janium they feed the specialized indexes —keyword search— and the authority file —alphabetical browse—. They are not a summary for the model: they are the index entries.
Article 50.2
Duty under Regulation (EU) 2024/1689, in force since 2 August 2026: anyone who places on the market an AI system that generates or manipulates text must mark that output in a machine-readable format so it can be detected as generated or manipulated, to the extent technically feasible. Collect marks cataloguing output that falls under that rule.
Authority control
The mechanism that ensures the same person, entity or subject is always recorded in the same form, checking it against reference lists —VIAF, ISNI, the Getty vocabularies, or the institution’s own catalog—. It keeps the same author from entering the index under three different spellings.
Systems
Janium
A web information-management system for libraries, archives and digital repositories. It takes records in standard formats (MARC21, Dublin Core, ISAD-G and others), runs the public catalog, circulation and authority control, and can index the full text of digital objects. It can run on the institution’s own servers or as a cloud service (JaniumNet). Collect produces the records; Janium is the catalog that searches them and circulates them. Product page: janium.com/janium.
Is there a term you would like to see here? Write to us at info@janium.com.