Czech non-breaking spaces for Typst, done by the rules.
Lines never end where Czech typography forbids it, and jump-to-source keeps working.

Quick start · What it binds · Options · Legal texts · Performance · How it works

In Czech typography, „vlna" (literally “wave”) is a non-breaking space, the tilde ~ in Typst and LaTeX. This package inserts it for you, so a line never ends after a one-letter preposition, inside a number or a date, between a number and its unit, inside a compound abbreviation, and in every other place the guideline of the Institute of the Czech Language (ÚJČ AV ČR, ?id=880) rules out. The feature set mirrors luavlna (Michal Hoftich), and every rule group has its own switch.

Quick start

#import "@preview/vlna:0.4.1": apply-vlna
#show: apply-vlna

That is all: the defaults follow the ÚJČ guideline. To see exactly what the package glued, outline every group:

#show: apply-vlna.with(debug: true)

What it binds

Groups chain: k s bratrem, o 2 500 Kč, od 21. 6. and plk. gšt. doc. Mgr. et Mgr. Ing. Libor Kutěj each stay one block. Code in raw, inline or block, is never touched.

Options

Every rule group can be switched off, or on, independently:

• units replaces the default list. Extend it instead: units: default-units + ("µm", "dpt") (default-units is exported).
• number-word applies “číslo + název počítaného jevu” literally and removes many break opportunities, which is why it is off by default.
• The uppercase branch of abbreviations allows 2–4 letters per part, so ALL-CAPS headings are not glued word by word.
• links inserts zero-width spaces (U+200B) into the displayed address, the technique the ÚJČ note itself recommends. The link target (dest) is untouched and stays clickable.

Legal texts

#show: apply-vlna.with(legal: true)

Off by default. Keeps the identifiers of Czech court filings and statutes in one piece:

The identifier is a block of its own inside č. j. KRPA-…: when the whole run does not fit on a line, the break falls after č. j., never inside the identifier. In legal mode věta and věty also bind in ordinary sentences, and hyphenated codes such as COVID-19 are glued.

Configuration guide

// Bind only one-letter words; `na`, `po`, `je`, `se`, … may end a line:
#show: apply-vlna.with(bind-two-letter-words: false)

// Bind exactly these words and nothing else (case-sensitive).
// A string is a set of single letters: this binds one-letter prepositions,
// their uppercase forms and uppercase A/I, while lowercase `a`/`i` may end
// a line:
#show: apply-vlna.with(short-words: "vksuozVKSUOZAI")

// An array lists whole words, so two-letter words can be included:
#show: apply-vlna.with(short-words: ("v", "k", "s", "z", "a", "i", "na", "Na"))

#set text(lang: "cs")             // required: the default text.lang is "en"
#show: apply-vlna.with(short-words: (
  cs: "AIiVvOoUuSsZzKk",          // luavlna's default set for Czech
  sk: "AIiVvOoUuSsZzKk",
))

#show: apply-vlna.with(short-words: (cs: "AIiVvOoUuSsZzKk"), lang: "cs")

#show: apply-vlna.with(compound-initials: ("Ch",))  // default
#show: apply-vlna.with(initials: false)             // turn initials off

#import "@preview/vlna:0.4.1": apply-vlna, vlna-on, vlna-off, vlna-debug-on, vlna-debug-off
#show: apply-vlna

Tady se váže. #vlna-off() Tady ne. #vlna-on() A tady zase.
#vlna-debug-on() Orámuje zásahy odsud dál. #vlna-debug-off()

Performance

Every match becomes a #box, so the package cannot be free, but it only pays for what the document needs. Measured on a 63-page Czech document with every rule group firing: typst 0.15.1, all series compiled in parallel, each pinned to its own physical core (2 threads), at least 12 runs per series. Harness, chart and raw dataset are in examples/benchmark.

v0.4.1 compiles 27% faster than v0.3.0 at p95, matches v0.4.0 (the new citation abbreviations cost nothing measurable), and legal: true adds 0.42 s, 27% on top.

These figures come from the quick parallel harness (about a minute, 12–85 runs per series) and are not comparable to the 501-page serial figures of earlier releases. The raw rows are the noisiest (up to 3 of 12 runs stalled), so their ordering is within noise; the serial 0.4.0 measurement (22.16 s → 20.30 s p50 on 501 pages, see CHANGELOG.md) remains the reference for that comparison.

The raw rows are the price of two features. #vlna-off(), #vlna-on(), #vlna-debug-on() and skipping raw all depend on where a match sits in the document, which forces a state lookup at every match, and a single code block anywhere is enough to require it. A document that uses neither takes the cheap path automatically, with no configuration.

How it works

A show rule that rebuilds text (it.text.replace(" ", "~")) produces new content whose source position points at the rule itself, so clicking a letter in the PDF jumps into the package code instead of your document. vlna wraps the original matched content in #box(..) instead. The box cannot break internally, and the glyphs keep their original source span, so clicking them jumps to your document.

Limitations

• Inside a box, spaces do not stretch under justification and words are not hyphenated, so a very long word right after a preposition can produce a looser line. A glued group wider than the line wraps at the column edge.
• Matches cannot cross styling boundaries: in k *lesu* the k is not bound, because the bold word is a separate element.
• Initials are a heuristic. A sentence ending with a bare capital (… vitamin C. Nový výzkum) looks like an initial and gets bound too; use initials: false if that bothers you.
• Text copied from a PDF may contain the invisible U+200B characters that links inserted, and jump-to-source is not preserved for the rewritten URL text (it is for everything else).

Version history

Compile times are not like-for-like across versions: 0.1.1 and 0.2.0 implement fewer rules, so they have less work to do and still lose. Full notes are in CHANGELOG.md.

Development

python3 tests/breaks.py   # line-break regression test (needs typst, pdftotext)
typst compile --root . examples/stress-test.typ   # every rule, with debug frames

tests/breaks.py sweeps the page width so that a line break lands in every gap of each glued unit, and fails if a unit ever splits. The benchmark and the typographic equivalence check against the previous release are described in examples/benchmark.

License

MIT, © 2025–2026 iamanro.