arXiv versions

Most papers on arXiv carry more than one version — the v1 submitted to the archive and a few v2/v3 revisions as referees weigh in. Usually the latest version is what you want, but occasionally the appendix / proof / data-table you're chasing only survives in a specific earlier revision.

BiblioFetch supports three ways of referring to arXiv versions:

The first two are plain single-entry references — they go through the normal cascade. The @all and @v1,v3 forms are pseudo-refs: the run loop expands them into one FetchEntry per version before calling fetch_paper!, so every version ends up in the store as an independent entry with its own TOML record and its own cached PDF.

How @all discovers versions

The arXiv Atom API answers id_list=<id> with the entry's canonical URL in <id>, and that URL always carries the current vN suffix. BiblioFetch parses N out, then materialises 1..N — arXiv numbers versions sequentially, so no per-version probe is needed. One API call per @all pseudo-ref, regardless of version count.

Example

A job that grabs v1 and the latest published version of Vaswani et al.'s Attention is All You Need (for comparing changes between the initial preprint and the NeurIPS camera-ready):

using BiblioFetch

mktemp() do path, io
    write(
        io,
        """
[folder]
target = "papers"

[doi]
list = [
    "arxiv:1706.03762@v1,v7",
]
""",
    )
    close(io)
    job = BiblioFetch.load_job(path)
    (length(job.refs), job.refs[1].key)
end

(1, "arxiv:1706.03762@v1,v7")

load_job is offline — @-pseudo-refs survive into job.refs as single entries and only get expanded once BiblioFetch.run starts. length(job.refs) == 1 above confirms that.

Running from the shell

bibliofetch run examples/arxiv-versions-job.toml

Each expanded version writes a distinct .metadata/arxiv__<id>vN.toml record and a distinct <id>vN.pdf under the store's target. The entries' raw field keeps the original @… spec, so bibliofetch info still tells you which pseudo-ref produced each concrete version.

Running the probe from Julia

If you want the version list without running a full job — say, to script "fetch only even-numbered revisions" — call the API helper directly:

# using BiblioFetch
# versions = BiblioFetch.arxiv_list_versions("1706.03762")   # e.g. [1, 2, 3, 4, 5, 6, 7]
# even_vs  = filter(iseven, versions)

The helper accepts both bare ids ("1706.03762") and arxiv:-prefixed ids, and strips any vN suffix before querying — so arxiv_list_versions("arxiv:1706.03762v2") returns the same list as arxiv_list_versions("1706.03762").

When to use which form

• Single latest version — the default. No @ anywhere.
• Historical diff — use @v1,v<latest> to pull the two ends of a paper's revision history and diff their PDFs.
• Archival sweep — @all for a deep citation where the v1 and v2 differ enough to matter (common in tensor-network / DMRG papers where appendices get rewritten).
• Known specific revision — arxiv:<id>v<N> (no @) stays the simplest way to pin a single non-latest version.

Mixing forms in the same [doi].list is fine; each ref is parsed independently.