Every institution with a collection reaches the same bottleneck sooner or later: the material exists, but to find it and use it, it has to be described. A book, a file, a photograph, a recording, a contract — each one needs a catalog record in a standard format, produced by someone who knows how to catalog. That work takes time, and the people who do it well are scarce. Meanwhile the material piles up, often already digitized but undescribed, scattered across a cloud folder, a server and boxes of scans.
Janium Collect takes on the first part of that work. It watches the places where the institution keeps its documents, reads them with the help of a language model, and produces a catalog record in the format the institution uses. Whoever catalogs stops typing every record from scratch and moves to reviewing the ones the system proposes.
What it produces
Collect’s output is a structured record, not free text. Depending on the type of collection, that record comes out in the corresponding format:
- Dublin Core for a simple, interoperable description.
- MARC21 (and its MARCedit/RDA variant, and UNIMARC) for library catalogs.
- ISAD-G for archives, where provenance and hierarchy — fonds, section, series — matter more than the isolated item.
- CDWA for works of art and cultural objects, with the Getty vocabulary.
The institution doesn’t change its catalog or its descriptive standard to use Collect; the system adapts to the format it already uses.
What material it starts from
Collections are rarely homogeneous, and Collect reads several kinds of source with the same destination — a record:
- Text documents, in PDF or in one of the thirty-plus office and publishing formats it converts for processing.
- Images and scanned documents, with text recognition when needed, or by reading the image directly when text isn’t enough.
- Already-structured data — a spreadsheet, an inventory list — turned into records without recapturing them.
- Audio and video, transcribing the speech and, in video, taking keyframes per scene to describe the content.
The criterion that underpins trust
Collect doesn’t just transcribe: it identifies the material and completes the record with data the source didn’t carry —subjects, a date, sometimes the ISBN—, drawn from the model’s knowledge. That enrichment is the value. The risk is that a value that sounds reasonable but is wrong stays in the catalog looking correct.
Trust doesn’t rest on not enriching, but on knowing where each value comes from. What can be verified is verified and attributed: names and subjects are checked against authority files and controlled vocabularies and carry their source; what comes only from the model’s knowledge stays unmarked. Fields with a fixed rule —the classification number, for instance— are computed, not estimated. And each record goes through an evaluation that scores it and flags it for review when it falls below the threshold, so review focuses on what the system points to.
Where it fits
Collect produces the records; the catalog receives them. It integrates with the institution’s system — Janium or another — and can take back the corrections the cataloger makes, so that the source and the catalog don’t drift apart. The person who catalogs stays at the center of the process; what changes is where their time goes.
What it doesn’t do
Collect works with what the document contains: if the document doesn’t state a date or a measurement, the record doesn’t invent it. The quality of the output depends on that of the source material, and the system doesn’t replace the judgment of someone who knows the collection.
What changes is how much human attention is needed, and where. Each record goes through an evaluation that scores it, and the system stops those that fall below the minimum before they enter the catalog. That score lets each context calibrate how much to review: a library with catalogers can review record by record; a firm processing thousands of contracts, or an archive describing in ISAD-G at scale, has no one person per record and reviews by exception —addressing what the score flags and trusting the rest. The quality reached makes that mode workable where reviewing everything isn’t.
To continue the conversation
If your institution has digitized material waiting to be described, or a steady flow of documents to catalog, the bottleneck stops being typing every record and becomes reviewing them. If that’s your problem, write to us at info@janium.com; we’re interested in understanding what your collection looks like and how you describe it today.