strip-nikkud

name	strip-nikkud
description	Remove Hebrew vowel points (nikkud) and/or cantillation marks (te'amim) from Hebrew text. Use when the user asks for "unvocalized", "consonantal", "without nikkud", "ktiv male", "strip vowels", "remove te'amim/trope", or asks to clean Hebrew text fetched from Sefaria (which ships with full nikkud and often te'amim) for plain reading, search, or copy-paste.

Strip Nikkud

Two paths — pick based on environment:

• Offline regex (default) — pure Python, no network. Fast and deterministic.
• Dicta API — https://removenikud.dicta.org.il/'s public backend. Use when the user explicitly asks for "Dicta", or when handling unusual edge cases (Yiddish, mixed corpora) where you want their tool's behaviour rather than a raw regex.

When to invoke

• "Strip the nikkud from this verse"
• "Give me Genesis 1:1 without vowels"
• "Remove the trope marks"
• "Unvocalized Hebrew please"
• After a find-text call where the user wants the consonantal text only
• Preparing Hebrew text for a system that doesn't render nikkud well (search engines, plain-text exports, some fonts)

Path A — Offline regex (default)

Unicode reference

Range	Block	Default action
U+0591–U+05AF	Cantillation marks (te'amim / trope)	Remove
U+05B0–U+05BD	Nikkud (sheva, vowels, dagesh, meteg, etc.)	Remove
U+05BF	Rafe	Remove
U+05C1, U+05C2	Shin/sin dot	Remove
U+05C4, U+05C5	Upper/lower dot (rare, masoretic)	Remove
U+05C7	Qamats qatan	Remove

Preserve these — they are punctuation, not diacritics:

• U+05BE maqaf (־)
• U+05C0 paseq (׀)
• U+05C3 sof pasuk (׃)
• U+05C6 nun hafukha (׆)
• U+05F3, U+05F4 geresh, gershayim (׳, ״)

Procedure

1. Confirm scope — strip nikkud only, teamim only, or both (default for "unvocalized").

2. Run via python3 -c in Bash. Use explicit \u escapes — the literal-character form silently includes punctuation code points and strips them too. The character class below has deliberate gaps at U+05BE / U+05C0 / U+05C3 / U+05C6 to preserve maqaf, paseq, sof pasuk, and nun hafukha:

   import re
   
   # Te'amim (cantillation) only — U+0591 to U+05AF
   TEAMIM = r'[֑-֯]'
   
   # Nikkud only — vowels, dagesh, sheva, meteg, rafe, shin/sin dot, qamats qatan
   NIKKUD = r'[ְ-ׇֽֿׁׂׅׄ]'
   
   # Both, in one pass — note the gaps that preserve Hebrew punctuation
   BOTH   = r'[֑-ׇֽֿׁׂׅׄ]'
   
   def strip(text, mode='both'):
       pattern = {'nikkud': NIKKUD, 'teamim': TEAMIM, 'both': BOTH}[mode]
       return re.sub(pattern, '', text)
   

   Verified: applied to בְּרֵאשִׁ֖ית בָּרָ֣א אֱלֹהִ֑ים אֵ֥ת הַשָּׁמַ֖יִם וְאֵ֥ת הָאָֽרֶץ׃ (Genesis 1:1), BOTH mode yields בראשית ברא אלהים את השמים ואת הארץ׃ — sof pasuk preserved.

3. Normalise whitespace afterwards: re.sub(r' {2,}', ' ', text).strip().

4. Don't use NFD-decompose-and-drop-combining-marks. It over-strips and breaks mixed-script text. Always use the explicit ranges above.

Path B — Dicta API

Public, no auth, no key. The frontend at removenikud.dicta.org.il calls this directly:

curl -sS -X POST https://remove-nikud-2-0.loadbalancer2.dicta.org.il/api \
  -H 'Content-Type: application/json' \
  -d '{"task":"remove_nikud","data":"<HEBREW TEXT>"}'
# => {"results":"<stripped text>"}

Notes:

• Undocumented API. Treat it as best-effort — if the URL 404s, fall back to Path A.
• Be polite: one request per user action, no tight batch loops.
• Strips both nikkud and te'amim — no mode toggle exposed.

Edge cases

• Mixed-script input (Hebrew + English + transliteration) — the Path A regex only matches Hebrew code points, so Latin/Arabic/etc. characters pass through untouched. Safe to apply globally.
• Final letters (ך ם ן ף ץ) are separate code points from their non-final forms and are not affected — don't try to "fix" them.
• Yiddish text uses some of the same code points (e.g. U+05B7 patah, U+05BC dagesh) as semantic letters, not vowels. If the source is Yiddish, ask before stripping — the user may want to keep them, or prefer Path B which is trained on Hebrew specifically.
• Te'amim-only mode (Tanakh with vowels but no trope) is Path A only; Dicta's endpoint doesn't expose it.

Output

Return the cleaned text in a fenced code block so RTL rendering doesn't reflow the user's terminal. Cite which path was used.