Install · Use · Routes · How it works · Accuracy · Roadmap

A quality gate for Arabic text extracted from PDFs. harfix tells you, per page and per font, whether the text you are about to index is trustworthy, needs a cheap repair, needs re-extraction with a different tool, or must go to OCR.

Arabic PDFs fail silently. A page can render perfectly and still extract as the wrong letters (المدعى عليه → الددعى عليهي), as presentation-form code points that match no search query, as mirrored lines, or as pre-Unicode symbols (Âd‡;« ≤∞). Generic "broken text" detectors check for control characters and Latin mojibake, so they pass this text straight into your RAG index. harfix scores the letters themselves with a small Arabic character model and reports what is wrong and what to do about it.

harfix never modifies text. It only decides.

Install

pip install harfix        # or: uv add harfix

Python 3.10+. The only runtime dependency is pdfminer.six; the two character models (~300 KiB) ship with the package.

Use

Audit the text layer of PDFs:

$ harfix scan --pages 8-12 --fail-on ocr mukhtasar_khuqayr.pdf
mukhtasar_khuqayr.pdf
  pages: keep=1, ocr=4
  p9     ocr       legacy_encoding
  p10    ocr       legacy_encoding
  p11    ocr       legacy_encoding
  p12    ocr       legacy_encoding

That book was typeset in pre-Unicode Arabic fonts (AXt…, Adv…), so its text layer extracts as Latin gibberish. -v lists every font on every page; --json prints the full report; --fail-on repair|reextract|ocr turns the scan into a CI or pipeline gate (exit code 1).

Check text produced by any extractor (Docling, pdftotext, Marker, an OCR engine). Pages separated by form feeds (pdftotext's default) are judged one by one; other text in runs of whole lines holding about 500 Arabic letters, so a few bad pages are not averaged away:

$ pdftotext -f 1 -l 12 al_mughalatat.pdf - | harfix check
reextract: bidi_controls, lam_alef_transposition, presentation_forms
  pages: repair=5, reextract=2
  p3               repair    bidi_controls
  p4               repair    bidi_controls, presentation_forms
  p5               repair    bidi_controls, presentation_forms
  p6               reextract bidi_controls, presentation_forms, lam_alef_transposition
  p7               repair    bidi_controls
  p9               repair    bidi_controls, presentation_forms
  p11              reextract bidi_controls, presentation_forms, lam_alef_transposition

Page numbers count from the start of the input; blank pages are skipped.

Repair what is safe to repair. harfix repair changes only the pages check routes to repair, and only for the reasons found on them: it deletes bidi embedding marks and unmapped elongation glyphs, turns visual-order lines into logical order (numbers and Latin words keep their own order, brackets are mirrored), and decomposes presentation forms. Pages that need reextract or ocr are left untouched and listed:

$ pdftotext -f 1 -l 12 al_mughalatat.pdf - | harfix repair -o repaired.txt
repaired p3: bidi_controls (8 characters)
repaired p4: bidi_controls (62 characters)
repaired p4: presentation_forms (278 characters)
repaired p5: bidi_controls (172 characters)
repaired p5: presentation_forms (289 characters)
repaired p7: bidi_controls (8 characters)
repaired p9: bidi_controls (12 characters)
repaired p9: presentation_forms (66 characters)
not repaired p6: reextract (bidi_controls, presentation_forms, lam_alef_transposition)
not repaired p11: reextract (bidi_controls, presentation_forms, lam_alef_transposition)

Decode pre-Unicode fonts without OCR. harfix decode reads the glyphs of each page and decodes them with font maps learned from typed editions of other books in the same fonts (packaged: the AXt… / Adv… family). A page is decoded only when the map knows at least 80% of its glyphs, and its text is written out only when harfix's own check passes it (keep or repair); other pages stay empty, so page positions are kept:

$ harfix decode --pages 8-12 mukhtasar_khuqayr.pdf -o decoded.txt
not decoded p8: font map knows 0% of its glyphs
decoded p9
decoded p10
decoded p11
decoded p12
$ sed -n 3p decoded.txt
هذا المختصر أملاه الشيخ أبو بكر وير في أخر حياته بعد عام

Decoded text gets about 70-85% of words right, so it is meant for search and triage, not for quoting. --json prints every page with its share of known glyphs and its check.

From Python:

from harfix import assess_document, assess_text, decode_pdf, repair_text, scan_pdf

report = scan_pdf("book.pdf", pages=range(1, 11))
for page in report.pages:
    if page.route != "keep":
        print(page.page, page.route, page.reasons)

document = assess_document(extracted_text)  # extractor output, page by page
document.route, [(s.page, s.assessment.route) for s in document.segments]

fixed = repair_text(extracted_text)  # repaired text, a change log, unrepaired pages
fixed.text, fixed.changes, fixed.unrepaired

verdict = assess_text(snippet)  # one piece of text: checks reading order too

pages = decode_pdf("book.pdf")  # DecodedPage: text, known_share, assessment, trusted
[p.text for p in pages if p.trusted]

Routes and reasons

quran_glyph_codes: Quran text set in the King Fahd Complex page fonts (QCF_P…) stores one glyph ID per word under Presentation Forms-A code points. NFKC would turn those IDs into wrong words, so they are never sent to repair; take the verse from a canonical Quran text or OCR it. The rest of the same text is still judged on its own letters.

A page takes the most severe route among its font runs, so one broken heading font next to a healthy body font is reported precisely instead of being averaged away. A font whose runs are too short to judge on any one page (running heads, headings, verse numbers) is judged on its short runs pooled across the scanned pages; such font reports carry "scope": "document". Fonts are identified by their PDF name including the subset tag (ABCDEE+Lotus Linotype), because two subsets of one typeface can map glyphs differently.

scan (raw text layer) ignores reading order, which depends on the extractor. check (final extractor output) also reports order and spacing faults.

How it works

1. Text is reduced to base Arabic letters (presentation forms decomposed, diacritics, tatweel and decorative letter runs removed).
2. Each line's reading direction is decided by comparing the model's score forwards and backwards, so mirrored text is not mistaken for corruption.
3. A 4-gram character model with interpolated Witten-Bell smoothing scores the corrected letters. Wrong glyph-to-letter maps produce improbable letter sequences (high bits per character) or isolated very improbable letters (spikes).
4. A second model trained without spaces separates "the letters are wrong" from "the extractor broke the word spacing".
5. Character-class checks catch private-use glyphs, raw (cid:N) codes, presentation forms, Quran glyph-ID fonts, and pre-Unicode encodings without any model. Pre-Unicode fonts land either on Latin-1 and symbol characters or on ASCII, where they read as Latin "words" with almost no vowels or with digits, symbols and case flips inside (otsrqponmlkA{zvjihgf, h&c%G); English text has about 40% vowels and few such words.

Accuracy

Real books

Full numbers: reports/benchmark.md. 57 printed-edition PDFs with a text layer, linked from al-Maktaba al-Shamela (up to six per subject category), up to 40 evenly spaced pages each: 2,280 pages. Labels come from Shamela's typed transcriptions, not from harfix: a font run is broken when at most 35% of its words exist in those transcriptions (or an English word list) and clean from 75% up. No OCR is involved. Of 24 labelled runs checked by hand, 23 matched their label and one was borderline.

The books are split by a hash of the PDF into 32 dev books, used when changing rules, and 25 test books. On the test split:

More than half of the labelled pages have a broken text layer, mostly pre-Unicode fonts and fonts with no Unicode map. Two rules added after 0.1.0 from dev books raised test font-run recall from 92.4% to 98.0%: pre-Unicode fonts mapped onto ASCII digits and punctuation, and U+FFFD elongation glyphs inside otherwise intact words (now repair instead of ocr). Samples of two test books had been looked at before the split, so the test numbers are close to, but not strictly, held out.

Repair

Full numbers: reports/repair.md. On the same Shamela PDFs, every page harfix repair changed is scored by the share of its Arabic words a plain search index would match in Shamela's transcriptions (only diacritics and tatweel removed):

The bidi embedding marks pdftotext inserts sit on about half of its pages in this sample and stick to the neighbouring words, so they alone hide many words from a search index.

Decode

Full numbers: reports/font_decoder.md. A map learned from 597 pages of four AXt… / Adv… books, scored on three other books in the same fonts against Shamela's typed text of the same pages, words as multisets:

The packaged map is trained on all seven books.

Synthetic corruptions

Full numbers: reports/calibration.md. Thresholds were fitted on one held-out split of 462 public Arabic pages and measured on another, then raised (DOMAIN_FLOORS) after inspecting clean runs in Egyptian court and statute PDFs.

On the test split, windows of at least 80 Arabic letters:

Limitations

• The model is trained on about 300 pages of modern standard Arabic. Classical, dialectal, or heavily domain-specific text scores higher; calibrate before relying on it there.
• Precision is preferred over recall: partial letter swaps and light padding are missed part of the time.
• scan cannot tell a text layer without word spaces from one with wrong letters: both are reported as same_script_substitution.
• A pre-Unicode font is caught only once its runs hold about 15 Latin-looking words or 8 digit-and-symbol words; a font whose codes land on plain numbers is not caught.
• unmapped_elongation assumes U+FFFD inside a word that reads correctly without it was an elongation glyph. A lost letter whose word still reads as Arabic slips through.
• check splits text without form feeds at line ends, so a single very long line is judged as one piece.
• Scanned pages are routed to OCR but their image quality is not assessed.
• decode knows one family of pre-Unicode fonts (AXt… / Adv…, seven books). Verse text in its Adv.QAL603 font is not decoded, and fonts that extract as raw glyph IDs ((cid:N)) are numbered per PDF, so their maps cannot be learned once for all books.
• PDF parsing uses pdfminer.six: about 0.1–0.2 s per page on the documents above.

Roadmap

• More decode font families, including verse fonts and glyph-ID fonts (matched by glyph shape instead of code), and a comparison with a modern vision-language OCR model.
• Docling, LangChain and LlamaIndex integrations.
• A public benchmark of broken Arabic PDFs with labels per failure class.

Rebuilding the model and thresholds

uv sync --group model
uv run python scripts/build_model.py    # pinned Hugging Face revisions, reproducible output
uv run python scripts/fetch_shamela.py  # needs the Shamela desktop app's catalogue
uv run python scripts/calibrate.py --pages 1-40 --pdf corpus/shamela/*.pdf

# real-book benchmark
uv run python scripts/fetch_shamela.py --out corpus/shamela_text --per-category 6 --text-only
uv run python scripts/benchmark.py fetch-text   # Shamela transcriptions, about 80 minutes
uv run python scripts/benchmark.py run
uv run python scripts/benchmark.py repair       # search matches before/after repair

# decode font maps (typed editions are fetched on first use)
uv run python scripts/fetch_shamela.py --out corpus/shamela_more --per-category 25 \
    --max-mb 30 --text-only --seed 1
uv run --group decoder python scripts/train_font_maps.py --train 9 10 19 1291 \
    --test 18 1575 1576
uv run --group decoder python scripts/train_font_maps.py \
    --train 9 10 18 19 1291 1575 1576 --save src/harfix/data/font_maps.json.gz

The Shamela PDFs and transcriptions are third-party works: keep them local, never commit or redistribute them.

License

Apache-2.0. The packaged character statistics are derived from Apache-2.0 datasets by Misraj; see NOTICE.