parseland-lib

The parsing library behind parseland, plus its evaluation harness and dashboard. Code without evals isn't really done — so they live here.

Example

from parseland_lib.parse import parse_page
from parseland_lib.s3 import get_landing_page_from_r2

url = 'https://doi.org/10.1002/andp.19033150414'
lp = get_landing_page_from_r2(url)
response = parse_page(lp)
print(response)

Layout

parseland-lib/
├── parseland_lib/        Library source
├── tests/                Library tests
├── eval/                 Offline eval harness (Python)
│   ├── parseland_eval/       Harness package
│   ├── runs/                 Benchmark run JSON
│   └── Gold Standard For Parseland - Sheet1.{csv,json}
│                             100-row hand-annotated gold standard
└── dashboard/            Static dashboard (Vite + React + TS)

Two-minute quickstart

# 1. Eval harness
cd eval
/opt/homebrew/bin/python3.11 -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'
pip install -r ../requirements.txt
python -m parseland_eval fetch
python -m parseland_eval run --label baseline

# 2. Dashboard
cd ../dashboard
npm install
npm run dev                     # → http://localhost:5173

What gets measured

Per-field, at three strictnesses (see eval/README.md for the matrix):

Field	Strict	Soft	Fuzzy
Authors	last + first-initial	—	rapidfuzz ≥ 85
Affiliations	exact	normalized	token_set_ratio ≥ 85
Abstract	exact	Levenshtein / normalized	Levenshtein / raw
PDF URL	canonicalized exact	—	—

Aggregated per row, per publisher domain, and per failure mode (paywall, login, bot_check, broken_url, no_abstract, non_article, image_only, clean).

Baseline (2026-04-16)

Metric	Score
Authors F1 (soft)	33.1%
Affiliations F1 (fuzzy)	81.1%
Abstract Levenshtein	26.4%
Abstract present rate	27.0%
PDF URL accuracy	12.0%

Most Authors/Abstract loss comes from Elsevier linkinghub.elsevier.com redirects, Oxford "Thanks for visiting…" gates, and login-wall landing pages. See dashboard/ → heatmap + failure-mode bar.

Workflow

1. Establish baseline — python -m parseland_eval run --label baseline once.
2. Make a parser change in parseland_lib/publisher/parsers/….
3. Re-run — python -m parseland_eval run --label fix-elsevier-2026-04-16.
4. Compare — dashboard renders delta vs previous run; trend chart accumulates.