paperparse-client

A thin, dependency-free Python client for the paperparse service — turn a scientific-paper PDF into structured output (figures + captions + sub-panels, tables, equations, body markdown) from any backend, without pulling in paperparse's heavy dependencies or source.

pip install paperparse-client

from paperparse_client import PaperparseClient

pp = PaperparseClient(
    api_key="denario_live_...",          # a denario-web key (issue at denario-web.vercel.app)
    client_ref="skepthical:run_abc123",  # opaque per-run id — REQUIRED for B2B keys
)
res = pp.parse("paper.pdf", out_dir="run/")   # uploads, parses, downloads artifacts

print(res["title"], res["n_figures"], res["cost"]["cost_usd"])
for f in res["figures"]:
    print(f["label"], f["path"], f["caption"][:60])   # f["path"] = local PNG (out_dir set)

• Stdlib only. No requests, no paperparse code — just an HTTP wrapper.
• client_ref is required. An opaque identifier the caller generates (a run id or a user hash). It's used only for attribution (stamped on the service's trace); paperparse never inspects it. B2B API keys reject calls that omit it.
• Async + robust. Uses the service's job flow (upload → run → poll), so a long parse survives a dropped connection.
• out_dir downloads everything locally — document.md, figures.json, and each crop — and adds a local path to every figure/panel/table/equation in the result.

The result shape is documented in the paperparse repo's CONTRACT.md (versioned via schema_version). Upload is subject to the service's 32 MB request cap (papers are virtually always under it).