Skip to content

jnu/tiny-trie

Repository files navigation

tiny-trie

Trie / DAWG implementation for JavaScript.

Build Status

About

Construct a trie out of a list of words for efficient searches. The trie provides a #freeze method to dedupe its suffixes, turning it into a directed acyclic word graph (DAWG).

More excitingly, there is a Trie#encode method which outputs the trie in a succinct binary format. (The trie does not need to be DAWG-ified in order to encode it, but it usually makes sense to do so.) The binary format is output in Base-64 and can be transmitted as JSON.

To use an encoded trie, there is a PackedTrie class. This class can make queries against the trie without ever having to parse the binary file. This class has virtually no initialization cost and low memory overhead without sacrificing lookup speed.

There are no specific character or size constraints on the Trie input. Unicode input should work, provided you treat the encoded string as unicode (it will contain the unicode characters somewhere in it.)

Features

Both Trie and PackedTrie support test and search methods which support fuzzy-matching (i.e., wildcards) and prefix search.

Docs

See complete docs at https://jnu.github.io/tiny-trie/

Quick Usage

const words = ['spit', 'spat', 'spot', 'spits', 'spats', 'spots'];

const trie = new Trie();

words.forEach(word => trie.insert(word));

// test membership
trie.test('spit');
// -> true
trie.test('split');
// -> false
trie.search('sp*t', {wildcard: '*'});
// -> ['spit', 'spat', 'spot']
trie.search('spi', {prefix: true});
// -> ['spit', 'spits']

// finalize the trie, turning it into a DAWG
trie.freeze();

// encode the trie
let encoded = trie.encode();
// -> 'A4AAAAMEspiaotI0NmhqfPzcsQLwwrCCcBAQE'

// This string describes the DAWG in a concise binary format. This format can
// be interpreted by the `PackedTrie` class.
const smallDawg = new PackedTrie(encoded);

smallDawg.test('spit');
// -> true
smallDawg.test('split');
// -> false
smallDawg.search('sp*t', {wildcard: '*'});
// -> ['spit', 'spat', 'spot']
smallDawg.search('spi', {prefix: true});
// -> ['spit', 'spits']

Including in a project

Installed in node_modules:

import {TinyTrie} from 'tiny-trie';
import PackedTrie from 'tiny-trie/lib/PackedTrie';

The default module export also provides some convenience functional tools:

import tinyTrie from 'tiny-trie';

tinyTrie.createSync(['foo', 'bar']);
// equivalent to:
//  > var t = new Trie();
//  > ['foo', 'bar'].forEach(word => t.insert(word));

tinyTrie.createFrozenSync(['foo', 'bar']);
// equivalent to:
//  > var t = new Trie();
//  > ['foo', 'bar'].forEach(word => t.insert(word));
//  > t.freeze();

Standalone files

Bundled, ES5-compatible equivalents to the above are in ./dist.

// tiny-trie[.min].js
TinyTrie.Trie
TinyTrie.createSync
TinyTrie.createFrozenSync

// packed-trie[.min].js
PackedTrie

Benchmarks

Quick benchmarks with the initial implementation on an MBP, node v5.0.0.

Using dictionary.txt, a Scrabble dictionary with 178,692 words.

var words = fs.readFileSync('./dictionary.txt', 'utf8').split('\n');

Speed

Gives an idea roughly how long things take.

> var trie = TinyTrie.createSync(words);
// 846 milliseconds

> trie.test(...);
// avg: 0.05 milliseconds

> trie.freeze();
// 124 seconds

> var encoded = trie.encode();
// 936 milliseconds

> var packed = new PackedTrie(encoded);
// 0.06 milliseconds (compare `new Set(words)`, which takes about 1s)

> packed.test(...);
// avg: 0.05 milliseconds (not significantly different from the unpacked trie!)

The init time of almost 1s is not acceptable for a client-side application. The goal of running Trie#freeze(); Trie#encode(); at build time is to produce a packed version of the DAWG that has virtually no init time - and it can still be queried directly, with speeds approaching the full Trie's very fast 50 microsecond times.

Memory

> words.join('').length
// 1584476 (bytes)

> encoded.length
// 698518 (bytes)

> encoded.length / words.join('').length
// 0.44085110787414894

The encoded trie uses just 44% of the bytes as the full dictionary. Gzipping gives a trie of 483kb, compared with 616kb for the dictionary.

TODO

  • Real benchmarks, comparison with other implementations

  • Optimize in PackedTrie - reduce size, increase perf. Node order could probably be revised to shrink pointer field width.

  • Spec out limitations on encoding inputs