Skip to content

paperus v0.1.1

Choose a tag to compare

@github-actions github-actions released this 13 Nov 13:58
· 80 commits to main since this release

"Select all" now actually selects all, and reading files no longer waits while arXiv does its three-second breathing exercises.

  • Reading files no longer waits behind lookups: reference lists,
    open-access copies, journal versions and authors are looked up in
    queues of their own, one per service (arXiv, Crossref, OpenAlex; under
    Jobs as one card, "Looking things up"), so files are read at full speed
    instead of between the pauses arXiv asks for (after a large import,
    about 10 files a minute), and no service's lookups wait behind
    another's. Lookups waiting at the update move there too, and "Reading
    files" counts only files.
  • Reference lists of journals' papers from Crossref first: on our
    papers as complete as arXiv's LaTeX sources, more often with DOIs, and
    without arXiv's 3 s pause and download; preprints still from arXiv.
  • "Last active" instead of "last login" among the users: any use of
    paperus (website, app, extension, assistant), not only logging in,
    which happens every few weeks.
  • The jobs charts' times no longer overlap: full minutes only.
  • Files in the trash (or of a paper in it) get no OCR and no vectors until
    restored, so trashing a large import of scans right away frees the ML
    service; a file deleted while its scans are read ends that quietly
    instead of with an error.
  • Imports say where papers were already: besides "already here",
    how many are in other libraries one sees ("1 already in Lab"). A paper
    in a group one may only read, with files the group's copy lacks, is
    added as one's own with them (before, its files were dropped without a
    word). The bell tells when an import is done (it may take long), with
    its counts.
  • Checkboxes on a computer: every row's box is always there, and the
    list's own box above them selects all of its papers (a dash when some
    are picked; a click then clears). On a phone, Select as before.
  • Select all in the selection's bar ("Select all 1,429"): every paper
    the list's filters match, also those not loaded yet, changed by the
    server in one go (up to 50,000). Before, only the rows loaded so far
    could be picked, and changing more than 1,000 picked papers failed; now
    they are sent in parts.
  • Whose papers the library lists, at hand: a menu beside Sort and
    Filter chooses All, Only mine or one of one's groups (for people in a
    group).
  • arXiv's versions of a preprint are versions of the paper: a PDF
    named with its arXiv version ("2401.12345v3.pdf") goes to the paper's
    preprint labelled "v3" (made if there is none), instead of becoming a
    second main file; the version cited is the published one, else arXiv's
    newest. Uploads, the browser extension, open-access copies and imports
    all count; existing preprints whose PDF names its version get the label
    on the update.
  • A paper's attachments in one row: code, data, supplementary files
    and slides, each chip saying what it is (before, links showed their long
    names and supplementary files hid under "Details and files"); "+
    Attachment" adds a link or uploads a file in that role.
  • The search models in a table, with what they take in memory: the
    ML service measures each model once loaded (RAM, or the GPU's memory),
    and the settings show it with the scores (English and multilingual),
    size and release date, the speed here, and whether the models in use fit
    together, with a warning when they do not. A model that does not fit is
    loaded after the idle ones are unloaded; on the GPU a reranker or OCR
    model then runs on the CPU.
  • Lists (library, collections, Recent) no longer stop loading at times
    while scrolled to their end, with only "Show more".
  • A long group name wraps on a phone instead of pushing the group's
    buttons off the screen.
  • Recommended models, for a CPU and for a GPU: choosing a model, the
    administration marks the ones to use where the ML service runs (setup
    too): arctic-embed-m-v2 for search and PP-OCRv6 for scans everywhere,
    the bge-reranker-v2-m3 reranker on a GPU. On 200 test questions
    arctic-embed-m-v2 was the best search model, and without a reranker as
    good as any with one, so a CPU, which does not rerank by default, loses
    little.