A PDF text extraction library that gets the hard parts right.
633eba61b1
- Create tests/fixtures/classifier/ with 200 synthetic PDFs:
- 50 invoices with bill-to/ship-to, item tables, totals
- 50 scientific papers with abstracts, sections, references
- 50 contracts with clauses, legal terminology, signatures
- 50 misc documents (8 receipts, 8 forms, 7 bank statements,
7 slide decks, 7 legal filings, 6 book excerpts, 7 magazines)
- Add MANIFEST.tsv mapping each document to its expected type
with source URL and license (all MIT-0 synthetic data)
- Add scripts/generate_test_corpus.py to regenerate the corpus
using reportlab for PDF generation
- Add tests/test_classifier_corpus.rs with validation harness:
- test_corpus_manifest_validity: verifies manifest structure
and file existence (PASSES)
- test_classifier_corpus_accuracy: will validate precision/
recall/F1 when classifier is implemented (SKIP for now)
- test_classifier_reproducibility: will verify deterministic
classification (SKIP for now)
- Add tests/fixtures/classifier/README.md documenting corpus
structure, generation process, and acceptance criteria
Total corpus size: ~0.4 MB (each PDF < 5 KB)
Acceptance criteria (from plan.md Phase 5.6):
- Per-class precision and recall >= 0.85
- Macro-F1 >= 0.88
- Reproducibility: identical output for same document
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
|
||
|---|---|---|
| docs | ||
| notes | ||
| scripts | ||
| src | ||
| tests | ||
| Cargo.lock | ||
| Cargo.toml | ||
| README.md | ||
pdftract
A PDF text extraction library that gets the hard parts right.
What it does
- Correct reading order — layout regions are segmented and sequenced before text is emitted, handling multi-column pages, sidebars, footnotes, and mixed-layout documents without relying on PDF operator order
- Font encoding recovery — when
ToUnicodeCMaps are absent, wrong, or incomplete, pdftract works through a layered recovery pipeline: glyph name lookup via the Adobe Glyph List, font fingerprinting against known metrics and embedded checksums, and glyph outline shape matching - Structure tree extraction — PDF/UA and PDF/A documents encode their logical structure (headings, paragraphs, lists, tables, reading order) in a
StructTree; pdftract reads this directly when present, producing accurate semantic output at no extra cost - Per-page hybrid routing — each page is independently classified and routed to the appropriate pipeline: vector text extraction, full OCR, or assisted OCR where vector hints improve raster accuracy
- Structured output with provenance — the primary output is JSON carrying per-span bounding boxes, font name, size, and confidence score alongside the extracted text, not a flat string dump
Output
{
"pages": [
{
"page": 1,
"blocks": [
{ "kind": "heading", "text": "Introduction", "bbox": [72, 680, 400, 700] },
{ "kind": "paragraph", "text": "...", "bbox": [72, 640, 540, 670] }
],
"spans": [
{ "text": "Introduction", "bbox": [72, 680, 400, 700], "font": "Times-Bold", "size": 14.0, "confidence": 0.99 }
]
}
],
"metadata": { "title": "...", "author": "...", "page_count": 10 }
}
Usage
pdftract extract invoice.pdf # structured JSON to stdout
pdftract extract invoice.pdf --text # plain text to stdout
pdftract extract invoice.pdf --output out.json
pdftract serve --port 8080 # HTTP service: POST /extract
Architecture
Rust core with PyO3 Python bindings and a CLI binary. The same binary runs as a command-line tool or as an HTTP microservice — the container deployment is just pdftract serve.
See docs/research/ for technical deep-dives into the PDF specification, font encoding, glyph Unicode recovery, and tagged PDF structure. See docs/notes/ for SDK invocation examples in Python, Node.js, Go, Ruby, Java, Rust, and Bash.
Status
Early development. See docs/plan/ for the implementation roadmap.