Do string operations based on graphemes instead of codepoints or bytes.
You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Brian Hicks 2973e72f86 bump to 1.0.2 4 years ago
benchmarks get rid of elm/parser 4 years ago
data add grapheme break tests from Unicode 4 years ago
script get rid of elm/parser 4 years ago
size-benchmarks get rid of elm/parser 4 years ago
src/String generate compact test failures on classes 4 years ago
tests generate compact test failures on classes 4 years ago
.gitignore add doc test 4 years ago
.travis.yml cache /nix again 4 years ago bump to 1.0.1 4 years ago add code of conduct 4 years ago Create 4 years ago
LICENSE add LICENSE 4 years ago
Makefile add doc test 4 years ago clarify grapheme explanation in README 4 years ago don't regenerate everything every time 4 years ago
elm.json bump to 1.0.2 4 years ago
shell.nix start making compatibility layer 4 years ago


Do string operations based on graphemes instead of codepoints or bytes. Compare:

import String.Graphemes

String.toList "🦸🏽‍♂️" --> [ '🦸', '🏽', '\u{200D}', '♂', '\u{FE0F}' ]

String.Graphemes.toList "🦸🏽‍♂️" --> [ "🦸🏽‍♂️" ]

What's going on here? Graphemes? What are those?

Unicode defines a system for encoding characters as numbers. These numbers are called codepoints! For example, a is codepoint 97, usually written in hex like 0x0061. There is a huge range of possible codepoints (from 0x0000 to 0x10FFFF), although not all of these match a symbol.

Codepoints are more complex than numbers, though: for a variety of reasons, a codepoint is encoded using 7 bits instead of 8. That means that we can't use regular 32-bit integers to represent them!

We do this partially for historical compatibility with ASCII, and partially to save space. For example, you can encode a (0x0061) in 1 byte, but 🦸 (0x1F9B8) takes four. If they didn't vary in length, you would have to pad out a with 3 bytes worth of zeros just to support both in the same string!

There's another layer of optimization, though! Imagine if you had to store a separate character for each accent mark like a, à, ā, ä, and á. You'd have a lot of characters on your hands, even before considering capital and lowercase letters! Plus, some languages use multiple accents for some characters! The combinations get ridiculous really fast, but we only have 1,114,111 (0x10FFFF) possible codepoints! So what we do is hardcode some combinations (like ä) for efficiency, but make separate codepoints for accents and let the software figure out how to combine them. These are called diacritic marks. So in addition to the hardcoded ä, you can put a and ¨ together to get the same thing. You can do this with more-or-less whatever characters and marks you want.

If you get really wild, you end up with z̴̙͒ả̴̫̼̫̀̅ĺ̴̔̿͜g̷̨͇͉̊͐̚o̶̳̣̯͌̓ text!

This raises another problem, though… if I have ä, I think of that as a single character, not two. But if I've encoded it as two codepoints and ask for the string length, it may tell me I have two characters! We deal with that using our final level: the grapheme.

A grapheme is what you'd intuitively think of as "a character" in a writing system. Whenever you combining codepoints you're working with graphemes. This applies to diacritic marks, as we've already explained, and tons of writing systems use graphemes: Hangul, Devanagari, Thai, and Tamil among others! But it also applies to emoji! For example: 🦸🏽‍♂️ is composed of 🦸 + 🏽 + zero-width joiner (200D) + ♂ + variation selector 16 (FE0F). You tend to think of 🦸🏽‍♂️ as a single character—a very definite expression which can't really be broken up into constituent parts. That means it's a grapheme!

But, final subtlety: if you used 🦸 by itself it's a grapheme too. The point is not "what codepoints are there?", it's "what is the smallest useful unit when expressing meaning?"

So what?

The above means that when we ask questions like "how long is this string?" or "what is the first character here?" we sometimes mix three levels:

  1. the byte level. Operations like String.length and String.left operate here (or, more specifically, they operate at the UTF-16 level, which assumes that codepoints are two bytes wide.) You should probably never operate here when working with String in Elm. It will result in subtle bugs and corrupt data! If you know you're working at the byte level, use elm/bytes instead.

  2. the codepoint level. Here, our base superhero emoji is only one character, but our skin tone and gender (🦸🏽‍♂️) take more, as discussed. This particular combination happens to be 17 bytes but only 5 codepoints. Operations like String.foldl operate here (so you can safely measure codepoint length with operations like String.foldl (\_ len -> len + 1) 0 "whatever string".) You should operate here if you're implementing higher-level operations on the codepoints, like grapheme segmentation (hi!) or normalization.

  3. the grapheme level. Despite being 5 codepoints, 🦸🏽‍♂ is only one grapheme. Operations like String.Graphemes.toList operate here. You should operate here if you're working with unicode text in ways meaningful to a user.

To underscore, if you're modifying text that the user has entered, work at the grapheme level. This reduces the possibility of errors and increases the possibility that your program will "do the right thing."

Still not convinced? Here are some practical reasons you should work at the grapheme level in the browser:

  • If you operate at the byte level, you will split multi-byte characters into invalid unicode sequences. If you do the wrong thing with these sequences, you'll crash your user's browser. In fact, that's what started me writing this library! Everyone does it occasionally, but there are better ways.

  • If you operate at the character level, you will split skin tones and genders off of people emoji, split flags into country codes, and move diacritic marks around. Your user entered this text precisely in these cases, don't lose their meaning!

  • Think your app doesn't need those pesky diacritic marks? Think again! They're crucial to understanding in a lot of languages! For example, in Spanish, papa (potato) is different than papá (father.) Don't make your users call their dad a potato!

Frequently Asked Questions

What spec does this package implement?

The Grapheme Cluster Boundaries section of UAX #29.

Does this package correctly reverse strings with diacritics?

Yes! It reverses the order of the graphemes, not the codepoints. This means that it does not move diacritics around and emoji are perfectly safe.

import String.Graphemes

-- äo without normalization
String.Graphemes.reverse "a\u{0308}o" --> "oa\u{0308}"

-- compare with String
String.reverse "a\u{0308}o" --> "o\u{0308}a"

Does this package do normalization?

No, and it probably never will. It's a whole 'nother spec in the Unicode standard which doesn't really fit in this package.

That said, it looks like you could implement it in a similar way as the internal String.Graphemes.Parser, so give it a go in a new package of your own!

(n.b. normalization in this case means turning "a\u{0308}" into "ä", usually for the purposes of improving equality checks.)

Does this package segment words or sentences?

No, and it probably never will. Segmenting words and sentences is locale- and implementation-dependent, so it's really hard to address them in a general way. Rather than introducing confusion ("it should segment here… why doesn't it?") we only segment graphemes.

That said, word and sentence segmentation rely on grapheme segmentation, so you're on the right track by asking this! UAX #29 has guidance here.

Why not "fix" elm/core's String instead of writing a new package?

The String module solves a different—but overlapping—set of problems. For example, you do not always want to work with graphemes: sometimes you need to be able to decompose into codepoints or operate at the byte level. As usual, it's all tradeoffs.

That said, if it eventually becomes obvious that merging into core would be a good thing we may do that. In that case, we would probably just keep equivalents of String.Graphemes.uncons and String.Graphemes.foldl. Everything else is implemented in terms of those two operations.

Why a drop-in replacement? / Why does the code refer to String functions so much?

Unless you've worked with unicode strings a lot, it can be tricky to know which level (bytes, codepoints, or graphemes) you're operating at with any given time. So instead of giving you the functions you might need, and leaving you to implement the rest on your own, we provide all of them and only change the ones where you'd run into trouble.

But not all of the functions in String need to be modified. In those cases, we just pass through to the String functions!

This way, you don't have to worry about it. You could potentially do import String.Graphemes as String in a module, fix the type errors, and all of a sudden all your string operations work with graphemes.

Climate Action

I want my open-source activities to support projects addressing the climate crisis (for example, projects in clean energy, public transit, reforestation, or sustainable agriculture.) If you are working on such a project, and find a bug or missing feature in any of my libraries, please let me know and I will treat your issue as high priority. I'd also be happy to support such projects in other ways. In particular, I've worked with Elm for a long time and would be happy to advise on your implementation.


This code in this project is licensed under the BSD 3-Clause license, located at LICENSE in the source.

The documentation strings in String.Graphemes are derived from those in elm/core's String, © 2019 Evan Czaplicki, and licensed under the BSD 3-Clause license.

The grapheme break property data used here are © 2019 Unicode®, Inc., and licensed under their terms of use.