Languages

Shiori grew up Japanese-only; since 0.2.0 it reads any language a language pack provides, with dead languages as first-class citizens. Japanese stays compiled in — its analyzer is real code — while every other language is data: a directory of files the app loads at runtime, no recompile, no restart. This page covers installing and switching languages, the Koine Greek pack, building packs from Wiktionary, and what changes in the app when the active language is not Japanese.

The one rule underneath it all: nothing mixes across languages. Library, reader, dictionary, reviews, statistics, and conversation practice all operate in the active language; a Greek λόγος and a Spanish sol can never collide with your Japanese words, and difficulty statistics never average across languages.

What a pack is

A pack is a folder (or zip) with a manifest.toml and the language's data files: a dictionary, a full-form morphology table (every inflected form mapped to its lemma and parse), a frequency list, graded vocabulary levels, and optionally pre-annotated texts ready to import. The app only consumes finished packs — the pack compiler (shiori-packc) is a separate developer tool, and the format itself is documented in Language-Packs if you want to build or author one.

Settings → Languages

Everything language-related lives on one settings page.

Activating

The Active language dropdown switches the whole app. On first activation a pack's reference data (dictionary, frequency, graded levels) imports into the database, scoped so it can never touch another language's data; after that, switching is instant.

Installed languages

Each installed language gets a card showing what its pack ships: the license, then the data files present — dictionary, full-form morphology, frequency list, graded levels (with the scheme's name), and any fonts the manifest declares. A pack that ships texts (the Koine Greek pack does) shows their count next to an Import into library button: one click adds the pre-annotated texts to that language's library, skipping any already imported. The button appears once the language is active.

Installing a pack

Three ways in, under Add a language:

Install from folder — pick an unzipped pack directory.
Install from zip — pick a pack zip.
Download URL — paste a URL, optionally with a SHA-256 checksum; the download is verified against it before installing.

Installs take effect immediately — the language appears in the dropdown with no restart. The low-tech route also works: Open packs folder reveals the data directory's packs/<code>/, and a pack directory dropped there by hand is picked up on the next start.

Removing a pack — and your data

Remove deletes the pack's files from the data directory, nothing more: your library, vocabulary, and review history for that language stay in the database and come back if the pack is reinstalled. The confirmation dialog says exactly this. The active language cannot be removed — switch to another one first.

Switching from the home page

The app opens on a home view whose header carries a Language dropdown and a Manage languages shortcut to the settings page, so switching never requires a trip through settings. Everything on the page follows the switch: cards due today (counted by your local midnight, with a time estimate from your measured review pace), the continue-reading card (progress, estimated time left at your reading speed, unknown words ahead, difficulty verdict), and the reading-activity calendar.

Koine Greek: the corpus-first pack

The first pack language is Koine Greek, and it takes the opposite approach to a runtime analyzer: the texts come pre-annotated. The pack's texts are converted from MorphGNT, so every token arrives carrying a hand-verified lemma, parse, and gloss — better than any analyzer could do, because humans checked each word. For the reader this means:

Interlinear glosses — the furigana slot doubles as a gloss layer: each word's English gloss is drawn over it, governed by the same fade modes as furigana (unknown-only, first-X occurrences per book).
Parse decoded to prose — clicking a word shows the parse of that occurrence, decoded: "verb · imperfect active indicative · 3rd person singular", not a cryptic tag.
Exact statistics — the pack grades against GNT frequency tiers (Core 50×+, 30×+, 20×+, 10×+, 5×+), which track the classic read-the-GNT vocabulary curricula. The Greek New Testament is a closed corpus, so coverage numbers are exact, not estimates.
Forgiving search — dictionary lookup ignores accents, breathings, and case, and the search box additionally accepts betacode and Greeklish: logos or lo/gos finds λόγος. See Dictionary & Kanji for the details.

Polytonic Greek currently renders through a wide-coverage system font; packs can declare fonts in their manifest, but per-pack font downloads are not wired up yet.

The Koine Greek pack is compiled by shiori-packc from MorphGNT and installs like any other pack — folder, zip, or URL. (The Build-from-Wiktionary list below also offers Koine Greek as a different, dictionary-driven pack built from Wiktionary's Ancient Greek data, without the pre-annotated GNT corpus. Both packs share the language code grc, so only one can be installed at a time — installing either replaces the other, while your library and review history survive the swap.)

Build from Wiktionary

Settings → Languages → Build from Wiktionary generalizes the Japanese first-run model to ~19 languages: Czech, Danish, Dutch, Finnish, French, German, Hungarian, Indonesian, Italian, Korean, Latin, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, and Koine Greek. Pick one, click Build, and the app downloads public data from its stable upstream URLs and compiles the pack locally — nothing is hosted or maintained by Shiori.

Two things are downloaded:

kaikki.org's Wiktextract dump for the language (CC BY-SA 4.0 & GFDL) — the dictionary and the full inflection tables.
hermitdave's FrequencyWords list (CC BY-SA 4.0) — subtitle-corpus frequency ranks, where a list exists. Latin and Koine Greek have none, and the row says so; the pack still builds, just without ranks.

Each row shows the approximate download size — dumps are large, hundreds of MB up to ~1 GB. The download streams to disk with progress, an interrupted transfer resumes with an HTTP range request instead of restarting, and a failed build keeps the downloaded file so a retry doesn't re-fetch; once the pack installs, the gigabyte-class dump is deleted to reclaim the space. Already-installed languages show a Rebuild button instead.

What the built pack includes:

Inflection tables as grammar — Wiktionary's per-word form tables are inverted into the full-form lookup table, so every conjugated or declined form resolves to its lemma, and each form's tags decode to prose in the reader ("hablaba → hablar · first person · singular · imperfect").
Register labels — senses keep their labels (colloquial, archaic, vulgar, …), mapped onto the usage-register display.
Usage examples attached to senses.
IPA pronunciation — shown with dictionary entries only when Settings → Reading → "Show IPA with dictionary entries" is enabled (off by default).
Lemmatized frequency and graded tiers — each surface form's subtitle mass folds onto its lemma, so a verb is ranked by all its conjugations, and Top 500 / 1k / 2k / 5k tiers light up the Statistics page's level section.

The list covers whitespace-tokenized scripts with rich Wiktionary inflection data. No-whitespace scripts (Chinese) need a segmentation engine — a Shiori release, not a pack.

How analysis works without an engine

Pack languages have no compiled analyzer, and mostly don't need one:

Pre-annotated texts carry their analysis with them — nothing runs at all.
Plain-text imports and chat resolve each token through the pack's full-form table: unambiguous forms get their lemma and parse directly.
Ambiguous forms resolve by corpus frequency when one candidate clearly outranks the rest — and the reader's word panel shows an "Ambiguous form" box listing every candidate with its decoded parse, so one click re-points that single occurrence. The manual override always beats the frequency vote.
Contractions are pack data — au, im, della stay one token but count as function words, and the panel shows the expansion (au = à + le) with each component clickable. French and Italian elision is handled the same way: l'eau tokenizes as l' + eau, while aujourd'hui stays whole.
Germanic compounds split against the pack's own dictionary when the whole word has no entry — Arbeitsmaschine → arbeit + maschine, linking elements understood — with each part clickable.
Forms missing from the table try suffix rules the pack builder learned from its own data, accepted only when the dictionary confirms the guessed lemma. Anything still unresolved keeps the surface form as its own lemma — safe, never wrong.

The same machinery powers dictionary search: an inflected query like suis resolves to every candidate lemma (être and suivre), and results build in tiers of match closeness with frequency ordering inside each tier — see Dictionary & Kanji.

Practice in a pack language

Conversation practice follows the active language, with the persona defined by the pack. A dead language makes no pretense of native speakers: the Koine Greek pack discloses the synthetic persona up front and judges your writing against attested usage in the period's texts rather than native intuition. Composition exercises and translation drills over sentences from your own reading work in every language.

Because a local model that handles Japanese fine may write terrible Koine, each language can pin its own LLM model: with a pack language active, Settings → AI shows a Model override field (blank means the provider's default). See AI & Chat.

The hosted catalog

A browse-and-install catalog — an offline-cached catalog.json with one-click SHA-256-verified installs, generated by shiori-packc catalog — is fully plumbed and tested, but no catalog is published yet, so the Languages page shows no browse section. Build-from-Wiktionary covers discovery for modern languages meanwhile, and install-from-URL handles anything hosted elsewhere. When a catalog is published, the browse UI will appear against the machinery that already exists.

Where pack data lives

Installed packs live in the data directory under packs/<code>/, and Build-from-Wiktionary caches its downloads under web-sources/ until the build succeeds. The database scopes everything by language — see Data & Interop for the full data-directory layout and what survives a pack removal.