Which document formats AI can catalog is not a question of file extensions: it is a question of holdings. An institution digitizes late, in pieces and with whatever is at hand. In the same folder sit the PDF of a report, the scan of a paper file, the photograph of a catalog card or a work, and a text file someone exported from a system that is no longer in use. Cataloging that, if each type demands a separate treatment, becomes prior work: open, recognize, and only then describe.
Janium Collect takes those files as they stand and a language model reads them. The destination is the same: a catalog record in the format the institution already uses. The cataloger stops preparing the material format by format and moves on to reviewing what the system proposes.
PDF is the container in which most digitized material arrives at an archive, a library or a documentation centre. Sometimes it was born digital: a report, a thesis, a bulletin. Sometimes it is paper put through a scanner: each sheet is an image inside the same file. For the cataloger the distinction matters —one can be searched by word, the other cannot— but both need a record, and both enter Collect as PDF.
What the system reads is the document, not the file name. From a report or a file it produces two outputs. The catalog record: title, author, date, subjects, producer, the unit’s dates. It completes what the source did not bring when the model can sustain it, and marks what it inferred.
And the document text in structured markdown: what the PDF said, with headings, lists and tables when they are kept, in a form that can be indexed and displayed. That markdown does not replace the record. It serves full-text search, consulting the content next to the catalog entry, and, if the institution needs it, a RAG layer over text already anchored to the record. How those layers are ordered is in Access points, full text and RAG.
A long PDF —a report of tens of pages, a yearbook, a thick file— is cataloged as one piece: the record describes that piece. The markdown keeps the body, to locate a phrase or read the text without going back to the original. An illegible or almost empty PDF produces a poor record and poor markdown: the quality of both outputs follows that of the material; it does not replace it.
The image
An image in a collection is not decoration on the record: it often is the document. The photograph of a museum label, of a title page, of a plan, of a page that never made it into a PDF, of a work. Collect reads it and proposes the record from what is seen.
That covers two situations that are better not mixed. When the image is the object —a historical photograph, a work, a plan— the output is a description of what is visible. When the image reproduces a document —a card, a page, a caption— what is read is the text or the data on that reproduction. In both cases the file is an image; the record that belongs is not the same.
The image does not declare what is not in it. A photo of a painting does not carry the accession number or the measurements; a photo of a catalog card does, if the card carries them. Combining what is seen with what is written is the subject of From the image and the catalog card to the record.
Plain text
Some holdings are no longer a facsimile but text: an export from an earlier system, inventory notes, a dump from a website, a file someone kept because it was all that remained of the description. Collect reads it as it reads a document: the content is the text, and from it comes the record.
The limit is that of the file itself. Text with a title, dates and names makes a useful record. A list of codes with no context makes little, and the system does not fabricate the context the file does not bring. What the source does not declare and the model cannot sustain is left unfilled.
What else comes in
A report or minutes in an office document —Word and the like— can also be cataloged. In many institutions that material lives alongside PDF and images: internal production that never went through a scanner. Collect takes it; the record comes out the same, in the catalog’s format.
Audio and video are another source type. A recording or a video also yields a record, and those cases have their own explanation: from a sound recording to a catalog record and from a video to linked records.
A spreadsheet used as an inventory —each line a work, a book, a unit— is not cataloged as if it were one more document. There the material already comes tabulated; what the column states is transferred into the record, and Collect completes what is missing when the sheet only identifies.
The record that comes out
It makes no difference whether the file is a PDF, a photograph or text: Collect proposes the entry in Dublin Core, MARC21, ISAD-G or CDWA, according to what the institution already uses. The catalog does not change standard because the holdings arrive mixed. From the PDF, in addition, the markdown of the body remains, to search it and consult it.
The cataloger receives records to review, not a closed batch. A sharp PDF and a blurry image do not produce the same certainty, and the system flags the record when it falls short. Mixed formats cease to be prior work; they remain, when the original is bad, a problem of the holdings’ quality.
Where the limits are
A format Collect does not cover is not cataloged. Coverage reaches what a digitized collection usually holds —PDF, image, text, office documents— and it is not universal. A little-used proprietary or very old format may fall outside; it is worth knowing before assuming it comes in.
Reading does not repair the original. An illegible scan, an underexposed photograph or truncated text do not become a good record because the model «completes». Completing is enriching what can be sustained; it is not standing in for a document that cannot be read.
To continue the conversation
If your institution has holdings split among PDF, images and text, and the bottleneck is treating each type as a separate case before cataloging, this is what Collect tries to ease. Write to us at info@janium.com; we are interested in how mixed your material is and in which format you describe it, because that changes what can be expected.