The workspace map

Verbora is one Cargo workspace of 19 production crates. Knowing its shape makes the rest of this site easier to navigate.

What each crate is for

Crate	Public surface
verbora-core	5 traits, StopWords, StopWordLanguage, the process-global stop-word list
verbora-tokenizers	WordTokenizer, SegmentTokenizer, SentenceTokenizer
verbora-distance	7 metrics, weighted cost sets, substring search, PreparedPattern
verbora-phonetics	12 encoders, PhoneticIndex, BeiderMorse, phoneticize_tokens
verbora-ngrams	ngrams, Padded, char_ngrams, CharNGrams
verbora-normalizers	nfc, nfd, nfkc, nfkd, remove_diacritics
verbora-inflectors	6 inflectors, Rule, CaseMode, Gender
verbora-trie	Trie, FrozenTrie, KeysWithPrefix, PrefixMatches
verbora-transliterators	transliterate_ja, transliterate_ja_into, transliterate_ja_normalized, Rewrite
verbora-wordnet	WordNet, Storage, Sense, Pointer, PrebuiltIndex, relation traversal
verbora-tfidf	TfIdf, Document, Analyzer, DocumentScore, TermScore
verbora-sentiment	SentimentAnalyzer, VocabularyKind, Vocabulary, Contributions
verbora-classifiers	BayesClassifier, LogisticRegressionClassifier, MaxEntClassifier
verbora-stemmers	Porter/Snowball language stemmers, Lancaster, Japanese and Indonesian
verbora-spellcheck	Spellcheck, FuzzyIndex, DeletionIndex, Correction, Neighbor
verbora-tagger	BrillTagger, Lexicon, RuleSet, Trainer, Evaluation
verbora-analyzers	analyze, SentenceAnalysis, SentenceType, TaggedWord, Role
verbora-language	script detection, optional language detectors, phonetic recommendations
verbora-util	abbreviations, stop-word re-exports, graphs, path trees, topological ordering

The dependency graph

Deliberately shallow and acyclic. Three crates reach sideways on purpose: verbora-phonetics tokenizes internally for tokenize_and_phoneticize, verbora-transliterators shares the normalizer's kana tables rather than duplicating them, and verbora-language composes the phonetic encoders with the transliterators to turn a detected language into a recommendation.

verbora-core ──┬── verbora-tokenizers ─── unicode-segmentation
               ├── verbora-phonetics ──── verbora-tokenizers, regex
               ├── verbora-stemmers ───── verbora-tokenizers
               ├── verbora-tfidf ──────── verbora-tokenizers, rustc-hash, serde
               └── verbora-util ───────── rustc-hash

verbora-ngrams           (no dependencies at all)
verbora-inflectors ────── regex
verbora-normalizers ───── unicode-normalization
verbora-distance ──────── rustc-hash
verbora-trie ──────────── smallvec
verbora-tagger ────────── rustc-hash
verbora-analyzers        (no dependencies at all)
verbora-wordnet ───────── memchr, rustc-hash
verbora-transliterators ─ verbora-normalizers
verbora-spellcheck ────── verbora-distance, rustc-hash
verbora-sentiment ─────── verbora-stemmers, verbora-tokenizers, rustc-hash
verbora-classifiers ───── verbora-stemmers, verbora-tokenizers, rustc-hash
verbora-language ──────── verbora-phonetics, verbora-transliterators

rayon is not drawn because it is optional in all fourteen crates that use it, and whatlang is optional in verbora-language — none of them is pulled in by a default dependency.

verbora-core reaches for one crate, rustc-hash, because StopWords::contains runs on every token of every document a stemmer filters. A leaf crate can be used in isolation without dragging in data assets or a regex engine it does not need: verbora-ngrams and verbora-analyzers ship an empty [dependencies] section, and verbora-distance reaches for nothing but rustc-hash, with no Unicode character database of any kind behind it.

The rest of the repository

site/                     this site (VitePress)
docs/                     internal engineering and design notes
tools/bench-data/         generates the inputs every benchmark harness reads
benches/data/             those generated inputs
benchmarks/competitive/   head-to-head suite against pinned third-party crates
crates/verbora-examples/  compiled documentation snippets (dev-only)

verbora-examples is not a crate you depend on: it exists so that the code on this site is real. Every non-trivial snippet here is extracted into a generated example in that package and compiled and run against the actual crates — see Documentation is part of the code.

Next

• Cargo features — the four opt-ins and what they cost.
• Features overview — what each subsystem does and when to reach for it.