jedarden/pdftract
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
pdftract/tests/fixtures/grep-corpus
History
jedarden 9a8252e92d feat(bf-20fot): extend download script to generate manifest entries 
- download-grep-corpus.sh now generates manifest.csv entries during download
- Implements SHA256 checksum computation via sha256sum
- Extracts page counts using pdfinfo
- Captures file sizes in bytes via stat
- Records source URLs for each downloaded file
- Appends entries to manifest.csv with all 6 required fields
- Handles incremental downloads (skips existing files, appends to manifest)
- Added grep-corpus-generate-manifest.sh helper for bootstrapping existing corpus
- Generated manifest entries for all 1,260 existing PDFs
- Verified checksum accuracy against actual files

Acceptance criteria:
- ✅ download-grep-corpus.sh writes manifest.csv entries for each downloaded PDF
- ✅ All 6 fields populated: filename, source_url, page_count, file_size, checksum, license
- ✅ Script handles incremental downloads (appends to existing manifest)

Closes bf-20fot

Verification: notes/bf-20fot.md
2026-07-05 12:26:55 -04:00
..
corpus feat(pdftract-5bzpg): implement pdftract-grep-1000 CI benchmark skeleton 2026-05-25 08:53:23 -04:00
manifest.csv feat(bf-20fot): extend download script to generate manifest entries 2026-07-05 12:26:55 -04:00
README.md feat(bf-5for4): design manifest schema for grep-corpus 2026-07-05 12:21:25 -04:00
regenerate.sh feat(pdftract-5bzpg): implement pdftract-grep-1000 CI benchmark skeleton 2026-05-25 08:53:23 -04:00

README.md

pdftract grep-corpus

Benchmark corpus for pdftract-grep-1000 CI benchmark.

Purpose

This corpus contains 1000 PDFs (~100 MB total) used to benchmark and validate the grep feature's performance and correctness.

Structure

tests/fixtures/grep-corpus/
├── corpus/              # Actual PDF files
├── manifest.csv         # File metadata and expected match counts
├── regenerate.sh        # Script to rebuild the corpus
└── README.md            # This file

Usage

Running the benchmark

cargo bench --bench grep_1000

Regenerating the corpus

cd tests/fixtures/grep-corpus
./regenerate.sh

Corpus Requirements

The corpus must satisfy:

  • Size: 1000 PDF files, ~100 MB total
  • Content: Mix of vector and scanned PDFs
  • License: Public domain or permissive (CC BY-SA, MIT, etc.)
  • Determinism: Regenerable from source (no manual uploads)

CI Gates

The benchmark enforces these gates on every PR:

  1. Throughput: ≥ 50 MB/s on 4-core CI machine
  2. vs pdfgrep: ≥ 2× faster
  3. vs pdftotext+ripgrep: ≥ 3× faster
  4. Regression: ≤ 10% vs historical main

Status

TODO: Populate corpus (blocks on 7.8.1-7.8.9 grep implementation).

Sources (TODO)

Potential corpus sources:

  • arXiv API (public domain metadata)
  • Wikipedia article exports (CC BY-SA)
  • Synthetic PDFs via pdfjoin

Manifest Format

filename,source_url,page_count,file_size,checksum,license
doc001.pdf,https://example.com/doc001.pdf,10,102400,abc123...,public-domain
doc002.pdf,https://arxiv.org/pdf/1234.5678.pdf,15,98304,def456...,cc-by-4.0
...

Fields

  • filename: Relative path from corpus/ directory (e.g., doc001.pdf)
  • source_url: URL where the PDF was downloaded from
  • page_count: Number of pages in the PDF
  • file_size: File size in bytes
  • checksum: SHA256 hash of the file contents (for integrity verification)
  • license: License identifier:
    • public-domain: Public domain
    • cc-by-4.0: Creative Commons Attribution 4.0
    • cc-by-sa-4.0: Creative Commons Attribution-ShareAlike 4.0
    • mit: MIT License
    • other: Other permissive license (document in source_url comments)