Nabu
by arvicco
README.md
# Nabu
[](https://github.com/arvicco/nabu/actions/workflows/ci.yml) [](https://doi.org/10.5281/zenodo.21361956)
**A personal, local, license-honest library of the ancient world — that your
AI tools can read.**
Project site: **[arvicco.github.io/nabu](https://arvicco.github.io/nabu)** —
the library, tools, and sources presented in full, without the README
compression.
Nabu pulls the world's openly licensed digital corpora of antiquity — Homer
and the Greek canon, the Latin classics, documentary papyri from Egypt, the
Sanskrit epics, cuneiform tablets, the Bible in fifteen parallel witnesses,
Beowulf — into one library on your own disk. Everything is plain files plus
SQLite: searchable by word or by dictionary lemma, citable to the exact verse
or tablet line, honest about every text's license, and rebuildable from
scratch at any time. And because it ships a read-only [MCP
server](docs/mcp.md), the AI assistants you already use can search, quote,
and cite the whole library — while structurally unable to change a letter
of it.
Named for the Mesopotamian god of scribes, patron of the tablet house and
divine custodian of Ashurbanipal's library. It is not a website and not a
reader app: it is a pipeline plus a database, operated from the command
line, designed to outlive the services it draws from.
As of **2026-10-13** the shelves hold **2,374,951 documents /
112,521,554 passages** in 207 language codes — from proto-cuneiform
tablets of the late 4th millennium BCE to Meiji-era Japanese. The
summer waves built the core: **EEBO-TCP Phase I** brought the
early-modern English print library whole (24.1M passages — the
corpus's second language) under the maintainer's written grant, the
**Old Mandarin desk** arrived as a connected series of rhyme books
(the Guangyun, Wang Renxu's Qieyun, the 'Phags-pa 蒙古字韻, the
中原音韻 of 1324), and the historical slice of the **Open Korean
Historical Corpus** more than doubled the document count in a day
(1,198,779 documents from the Samguk sagi of 1145 through the
Ilseongnok court diaries, each record carrying its own license).
The autumn arc then opened whole regions: the **Southeast Asia
desk** (the DHARMA epigraphic corpora with the first digital Old
Mon text anywhere, the kakawin library, Bagan's Old Burmese — then
1,252 **Balinese palm-leaf lontars** transcribed manuscript-direct,
Jawi-script Malay manuscripts, the Thai Collected Chronicles), the
**Central Asia corridor** (Khotanese, Sogdian, Bactrian, Chagatai,
the Secret History of the Mongols), the **Slavic shelves end to
end** (nine Wikisource cones, the 1488 Prague Bible, the Ipatiev
chronicle), a first **American wing** (colonial Zapotec testaments,
Guaman Poma's chronicle), and **South Asia in force**: the DHARMA
Tamil-country and Deccan epigraphy, Siddham's Gupta pillars, the
printed *South Indian Inscriptions* volumes, Project Madurai's
classical Tamil e-texts at 1.2M lines, the Gurmukhī scriptural
canon at line grain — and **TAMILEX**, whose per-word gold
wordsplit layer (262,093 rows) exists nowhere else for Old Tamil.
Behind TITUS's members-only gate, personal grants opened the
**Middle Iranian books** and the **whole Tocharian corpus** — all
4,250 West Tocharian documents of the Berlin, London and Paris
collections, plus two leaves Maue's handlist identifies as
Tumshuqese. **OpenITI's Classical Arabic** remains the largest
language at 33.3 million passages, ahead of early-modern English
(24.3M) and Literary Chinese (19.0M).
Plus **2,761,739 dictionary entries** across 566 dictionary
shelves and **22.8 million gold lemma annotations in 48 languages** (a
further 31.0 million ride an honestly labelled silver tier). (All numbers in this
README are read from the live catalog, never estimated.)
## Quickstart
```
git clone https://github.com/arvicco/nabu && cd nabu
bundle install # Ruby 3.3+; deliberately small dependency set
bin/nabu quickstart # the starter shelf: 4 sources, ~690 MB, minutes — then:
bin/nabu align "MARK 2.3" # one verse, seven witnesses
bin/nabu search --lemma λέγω # dictionary-form search: λέγουσι, εἶπας, εἰπεῖν…
bin/nabu define λόγος # the whole LSJ entry, citations resolved
```
A fresh checkout works with zero configuration. The long form —
prerequisites, honest sizes and timings, growing the library, MCP
registration — is the site's
**[Quickstart](https://arvicco.github.io/nabu/quickstart/)** page
(kept in-repo as [docs/quickstart.md](docs/quickstart.md)).
## Show me
Real commands, real output, pasted from live runs on 2026-07-11/12 (trims
marked with …).
One verse of Mark across the aligned witnesses — Greek, Latin, Gothic, Old
Church Slavonic, Old English, and more — each with its license label (this
run predates the Sahidic and Bohairic Coptic witnesses that joined
2026-07-13, making fifteen registered):
```
$ bin/nabu align "MARK 2.3"
MARK 2.3 — New Testament (parallel witnesses)
13 of 13 witnesses attest this ref
greek-nt — The Greek New Testament [grc] license: nc
urn:nabu:proiel:greek-nt:6563
καὶ ἔρχονται φέροντες πρὸς αὐτὸν παραλυτικὸν αἰρόμενον ὑπὸ τεσσάρων.
latin-nt — Jerome's Vulgate [lat] license: nc
urn:nabu:proiel:latin-nt:10368
et venerunt ferentes ad eum paralyticum qui a quattuor portabatur
gothic-nt — The Gothic Bible [got] license: nc
urn:nabu:proiel:gothic-nt:37435
jah qemun at imma usliþan bairandans, hafanana fram fidworim.
marianus — Codex Marianus [chu] license: nc
urn:nabu:proiel:marianus:36421
Ꙇ придѫ къ немоу носѧште ослабленъ жилами. носимъ четꙑрьми.
wscp — West-Saxon Gospels [ang] license: nc
urn:nabu:proiel:wscp:102359
& hi comon anne laman to him berende, þone feower men bæron.
WEB (English) — Mark [eng] license: open
urn:nabu:eng-web:mrk:2.3
Four people came, carrying a paralytic to him.
… (Armenian, SBLGNT, Clementine Vulgate, and the four CCMH OCS witnesses trimmed)
```
Look up the first word of Western literature — with the dictionary's
citations resolved to live passages in your own catalog:
```
$ bin/nabu define μῆνις
μῆνις — A Greek-English Lexicon (Liddell-Scott-Jones) [attribution] urn:nabu:dict:lsj:n67485
gloss: wrath
μῆνις, Dor. and Aeol. μᾶν-, ἡ, gen.
A. μήνιος Pl. R. 390e , later μήνιδος Ael. Fr. 80 , … —wrath; from Hom.
downwds. freq. of the wrath of the gods, Il. 5.34 , al., A. Ag. 701 (lyr.),
… but also, generally, of the wrath of Achilles, Il. 1.1 , al. …
resolved citations (in this corpus — nabu show <urn>):
Il. 1.1 → urn:cts:greekLit:tlg0012.tlg001.perseus-grc2:1.1
Il. 5.34 → urn:cts:greekLit:tlg0012.tlg001.perseus-grc2:5.34
A. Ag. 701 → urn:cts:greekLit:tlg0085.tlg005.1st1K-grc1:701
…
```
Search by dictionary form, not surface string — suppletion and all:
```
$ bin/nabu search --lemma λέγω --limit 3
urn:nabu:proiel:chron:108755 [grc] λέγω → ῥηθέντος (lay)
ὅ περ ἦν καὶ αἴτιον τοῦ μὴ ἐλθεῖν τὸν γενήσαντά με εἰς τὸν Μορέαν μετὰ τοῦ αὐθεντοπούλου κὺρ Θωμᾶ εἰ…
urn:nabu:proiel:chron:121080 [grc] λέγω → εἶπον, εἰπὲ (lay)
Πολλῶν οὖν λόγων δαπανηθέντων, τέλος ἐστάλησαν πρὸς τὸν ἄνθρωπον δύο τῶν κελλιωτῶν καὶ συντρόφων μου…
urn:nabu:proiel:chron:121083 [grc] λέγω → εἴπω (lay)
Ἐγὼ δὲ νὰ ἀκούω παρὰ μὲν τῶν, ὅτι καλή ἐστι, παρὰ δὲ τῶν, ὅτι οὐ καλή, διὰ τὶ νὰ μηδὲν εἴπω·
3 hits (exact lemma match; text is pristine)
```
Pull a random tablet off the cuneiform shelf:
```
$ bin/nabu show --random --source oracc
urn:nabu:oracc:rinap-rinap1:Q003443:1 [akk]
a-di {KUR}sa-u₂-e KUR-e ša ina {KUR}lab-na-na-ma it-tak-ki-pu-u₂-ni
document: urn:nabu:oracc:rinap-rinap1:Q003443 — Tiglath-pileser III 30
source: oracc license: open sequence: 0 revision: 1
provenance:
2026-07-10 18:36:28 +0200 loaded nabu-loader
```
— a royal inscription of Tiglath-pileser III, "as far as Mount Saue, which
abuts Lebanon."
## Who this is for
- **Classicists.** The Perseus Greek and Latin canons plus First1KGreek —
2,209 Greek and Latin editions with 872 aligned English translations
(`show <urn> --parallel` pairs Vergil line by line). TLG-style proximity
search, lemma-aware:
```
$ bin/nabu search λόγος --near θεός --window 5 --lang grc
urn:nabu:ddbdp:p.oxy:8:1151:18 [θεοσ] ην ο [λογοσ]. ← a papyrus amulet…
urn:cts:…:tlg0031.tlg004…:1.1 …και [θεοσ] ην ο [λογοσ]. ← …quoting John 1:1
```
- **Biblical scholars.** The New Testament in up to **fifteen registered
witnesses** (`align "MARK 2.3"` → Greek ×2, Latin ×2, Gothic, Armenian,
five OCS manuscript editions incl. Assemanianus and both Marianus
editions side by side, Old English, English, and — live since 2026-07-13
— Sahidic and Bohairic Coptic), the Old Testament on the Septuagint ↔
Vulgate ↔ English axis with the Greek/Hebrew Psalm numbering mapped
honestly (`align "PSA 22.1"` shows WEB's 23.1 labeled).
- **Slavists & textual critics.** The OCS canon complete — Marianus,
Zographensis, Assemanianus, Savvina kniga, Suprasliensis (folio-line
cited, hyphen-split words searchable whole) — plus Old East Slavic from
birchbark to Ruthenian chancery texts, and the ~1000 CE Freising
Manuscripts in three aligned transcription layers. `align REF --collate`
turns the aligned witnesses into an apparatus: a raw-token diff per
script family (the four Helsinki-ASCII CCMH codices collated against each
other; the Cyrillic Marianus set beside them, honestly not collated
because the fold cannot bridge the two transcription systems).
- **Comparativists.** The reconstruction shelf walks attested words to
their proto-forms and cognates, with corpus attestation counts:
```
$ bin/nabu etym богъ --lang chu
богъ [chu] → *bogъ [sla-pro] — gloss: god
← *bʰeh₂g- [ine-pro] — gloss: to divide, distribute, allot
reflexes: [grc] ἔφᾰγον, [sa] भक्ष (bhakṣá), …
```
Pure-ASCII input works (`etym bhewgh`); `--long` expands every reflex.
And `cognates` crosses that crosswalk with the alignment hub — verses
where the witnesses use reflexes of the same root, found blind:
```
$ bin/nabu cognates "LUKE 14.34" --langs got,chu
LUKE 14.34 *sḗh₂l [ine-pro · attribution]
chu соль — attested as солъ
got salt
```
The whole Gothic × OCS NT yields ~300 such verses across 30 roots in
under a second (hlaifs ~ хлѣбъ, malan ~ млѣти, menoþs ~ мѣсѧць), each
hit labeled with its meet shelf — a gem-pro meet for a Slavic word is
flagged reading matter: likely a borrowing, not common descent.
- **Indologists.** 780 GRETIL editions, 703k passages: Rāmāyaṇa, purāṇas,
kāvya, dharmaśāstra, the Ṛgveda with Vedic accents preserved; commentary
layers separately citable (kārikā vs. vṛtti).
- **Papyrologists.** 61k DDbDP documents, and fragment search that reads
like an edition: type the damaged line brackets and all, and `--fuzzy`
matches it INSIDE words (trigram index over papyri + ORACC + EDH,
sub-10 ms):
```
$ bin/nabu search --fuzzy ']ανδρα μοι εν['
urn:nabu:ddbdp:bgu:6:1470:ctr:6 [grc]
μαρτυροι. [ανδρα μοι εν]νεπε μουσα πολυτρο
1 hit (fuzzy substring; highlights are diacritic-folded)
fuzzy index covers: oracc, papyri-ddbdp
```
— BGU 6.1470, a Hellenistic writing exercise breaking off mid-word
through the Odyssey's opening line (…Μοῦσα πολύτρο[πον). *(Live output
of 2026-07-13; the EDH inscriptions joined the index's scope later that
day — it now covers 1.71M passages.)*
- **Epigraphists.** 81,881 Latin inscriptions from the Epigraphic Database
Heidelberg (CC BY-SA; a preservation snapshot of the archived upstream) —
81,416 of them dated (2026-07-14 census), with the library's first genre
facets: `search
manibus --type epitaph --province Britannia` composes the genre facets
with the date and place filters, and the stones are in the `--fuzzy` index. Since
2026-07-17 the epigraphic shelves speak Celtic too: 428 Gaulish
inscriptions (RIIG, with per-editor readings and French translations)
and ~500 Irish ogham stones in real Ogham codepoints with aligned
transliteration layers.
- **Celticists.** The CorPH corpus (ERC ChronHib) brings 7th–10th-century
Early Irish with gold lemmatization — the Annals of Ulster, Vita
Columbae, Blathmac, and the Milan/St Gall/Würzburg gloss corpora —
joined by two Old Irish UD treebanks, the Gaulish and ogham epigraphy
above, and Old/Middle Irish and Middle Welsh dictionary extracts on the
reference shelf.
- **Assyriologists.** 21,692 ORACC documents (CC0) across 33 projects —
12,781 tablets and inscriptions, the complete State Archives of Assyria
among them, plus 8,911 aligned English translations — with gold
lemmatization in
`search --lemma` and the running English translations aligned per line:
```
$ bin/nabu show urn:nabu:oracc:saao-saa01:P224395:o.1-o.3 --parallel
:o.1 akk a-na LUGAL EN-ia
:o.2 akk ARAD-ka {1}10-ha-ti
… eng To the king, my lord: Your servant Adda-hati. …
```
- **Medievalists.** The complete Anglo-Saxon Poetic Records (Beowulf cited
by its real line numbers: `show urn:nabu:aspr:A4.1:1` → *Hwæt! We
Gardena in geardagum*), the ISWOC treebank with West-Saxon Mark as an
alignment-hub witness, and Bosworth-Toller on the dictionary shelf —
`define aethele --lang ang` finds **æþele** through the æ/þ/ð folding.
- **Germanicists.** All three branches on one desk since 2026-07-22:
Wulfila's Gothic (gold-lemmatized in PROIEL), the Old English shelves,
Menotec's seven Old Norwegian treebanks with the Poetic Edda of Codex
Regius, the *Heliand* parsed to the token, ReM's 355k gold-annotated
Middle High German manuscript lines (long ſ folded, so `swert` finds
`ſwert`), and the runestones of Rundata — each inscription with its
transliteration, Old-West-Norse and runic-Swedish normalisations, and
English/Swedish translations as `--parallel` lanes.
- **Linguists & digital humanists.** 21.9M gold lemma rows in 45 languages
(2026-08-31 census) with morphology facets (`search --lemma cyning --morph case=gen --lang
ang`), distinctive-vocabulary profiles (`vocab urn:nabu:proiel:cic-off` →
officium, honestas, decorum), and `export --format jsonl` streaming the
corpus to your own tooling with license filters.
- **AI-tooling builders.** A hand-rolled, dependency-free MCP server over
stdio (`bin/nabu mcp`, `.mcp.json` ships in-repo) exposes fourteen read-only
tools — search, show, concord, align, define, etym, parallels, cognates,
links, place, char, signs, clarify, status — every passage carrying its license class, so a model
can quote *and* cite responsibly. See [docs/mcp.md](docs/mcp.md).
## What's on the shelves
A representative slice, not the whole registry (151 corpus sources as of
2026-09-02); each row carries the date its counts were read. The full
shelf map with research uses per shelf is
**[docs/library.md](docs/library.md)**. Rows below without a date are the
2026-07-14 census.
| Shelf | What's on it | Size | License |
|---|---|---|---|
| Classical Greek | Perseus: Homer, the tragedians, Herodotus, Plato, Galen… + 650 aligned English translations | 1,418 docs / 394,706 passages | CC BY-SA |
| Post-classical Greek | First1KGreek: Athenaeus, Philo, church fathers, Swete's Septuagint | 1,129 / 256,480 | CC BY-SA |
| Classical Latin | Perseus: Vergil, Ovid, Cicero, Livy, Tacitus… + 181 English translations | 534 / 391,799 | CC BY-SA |
| Documentary papyri | Papyri.info DDbDP: contracts, letters, tax receipts from a millennium of Egypt (Greek, Coptic, Latin, Arabic) | 61,414 / 921,611 | CC BY |
| Latin inscriptions | Epigraphic Database Heidelberg: epitaphs, dedications, milestones from the whole empire — 81,416 dated (2026-07-14 census), genre/province/material facets | 81,881 / 406,306 | CC BY-SA |
| Coptic | Coptic Scriptorium: the complete Sahidic + Bohairic NT, monastic and patristic prose — gold-lemmatized (233k rows) | 482 / 74,169 | CC BY per doc (source class nc) |
| Sanskrit | GRETIL: Rāmāyaṇa, purāṇas, kāvya, śāstra, Ṛgveda with Vedic accents | 780 / 703,068 | CC BY-NC-SA |
| Treebanks | PROIEL, TOROT, UD, ISWOC: gold lemma/morphology/syntax — parallel NT ×5, OCS→Middle Russian, Old English, Old Irish glosses | 80 / 178,278 | mostly CC BY-NC-SA |
| Cuneiform | ORACC ×33 projects: the complete State Archives of Assyria, royal inscriptions, lexical lists, proto-cuneiform — with 8,911 aligned English translations | 21,692 / 385,243 | CC0 |
| Biblical editions | Clementine Vulgate (73 books), SBL Greek NT, WEB English | 184 / 81,372 | PD / CC BY |
| Old English poetry | The complete ASPR: Beowulf, the Exeter Book, Dream of the Rood… | 349 / 30,550 | CC BY-SA |
| Slavic & Slovenian | CCMH OCS gospel codices, the ~1000 CE Freising Manuscripts, goo300k + IMP Early Modern Slovenian (1584–1899), the damaskini Balkan Slavic witnesses (15th–19th c., with English siblings) | 839 / 456,189 | CC BY (Freising BY-ND) |
| Celtic | CorPH Early Irish (gold-lemmatized: Annals of Ulster, the great gloss corpora), RIIG Gaulish inscriptions (with French siblings), the Ogham in 3D stones (real Ogham codepoints + transliteration layers) | 1,387 / 20,318 | CC BY / MIT (ogham `nc` pending clarification) |
| Chinese library | Kanripo classical Chinese + the CBETA Buddhist canon (Taishō + Xuzangjing) — 13.2M `lzh` passages (2026-07-22; the language now totals 18.7M with the Korean hanmun shelves), trad/simp/z-variant spellings folded to one search skeleton | 13.2M passages | mostly CC BY-NC / CC BY |
| Japanese reading desk | Aozora Bunko: the Japanese public-domain library, Meiji core with a classical tail — ruby readings as annotations, kyūjitai reachable via the reform fold (2026-07-22) | 17,195 / 2,991,807 | open (PD grant) |
| Germanic wave | Menotec Old Norwegian treebanks + Poetic Edda, the Old Saxon *Heliand* (HeliPaD), ReM Middle High German (355k gold manuscript lines), Rundata runic inscriptions in five text lanes (2026-07-22) | 31,057 / 409,947 | nc / CC BY / CC BY-SA / odbl |
| Southeast Asia desk | The DHARMA corpora (Old Khmer, Campā, Nusantara, Pyu — with the first digital Old Mon anywhere), the Bagan Old Burmese inscriptions, the kakawin critical editions, the Old Javanese Wordnet (2026-09-02) | 2,985 docs / 130,761 passages + 3,632 entries | CC BY / BY-SA |
| Korean desk | The Veritable Records of Joseon (sillok) in the original hanmun, the Goryeosa family, and since 2026-09-01 the OKHC historical slice: the Samguk sagi (1145), the Ilseongnok, the ITKC munjip mass, the AKS/Kyujanggak/NHM old-literature dbs | 1,198,779 / 3,542,662 (okhc alone) | per-record PD/KOGL/nc (source CC BY-NC 4.0) |
| Reference shelf | LSJ + Lewis & Short + Bosworth-Toller + Monier-Williams + Wiktionary OCS + ten Wiktionary reconstruction/Celtic shelves + the IE-CoR / LIV / de Vaan etymological witnesses + the five StarLing bases (Pokorny, PIET, **Vasmer**, Germanic, Baltic) + three Slovenian historical dictionaries incl. Pleteršnik + the Hebrew/Egyptian/Slovene desks and the Sino-Japanese lexicography (Unihan, KANJIDIC2/JMdict, HDIC, Guangyun) (`nabu define` / `etym`) | 1,310,763 entries / 56 shelves (2026-07-22) | CC BY-SA / CC BY / CC BY-NC-SA / grant |
The registry holds **215 corpus sources + 5 local shelves + 25 feature
modules** (the kind-split census, 2026-10-13 — `sources.yml`
distinguishes what a row IS: a corpus that mints catalog rows, an
owner-authored local memory shelf, or machinery like `osl` (the Oracc
Sign List behind `nabu signs`), `nabu-data` (Nabu's own published
datasets, consumed back) and `pedecerto`/`bridging` that fetches
reference data but mints no documents of its own). All corpus
sources are **wired** — adapter built and first sync verified —
except **KorBa** (the Baroque Polish corpus: adapter ready, the
IJP's download server answering 502 since early October); the
newest arrivals are the October South Asia slice (`tamilex`,
`sii-clean`, `madurai`, `shabados`), the whole members-area
**Tocharian B chain** behind the new credentials seam, the
SEA/Gujarat Wikisource cones (`ban`/`gu`/`ms`/`th`), and the TITUS
Middle Iranian books — behind them the **StarLing etymology
bases** (Pokorny, Vasmer, the branch comparative shelves and the
LEXSTAT wordlists), the
**Southeast Asia seven** (the five DHARMA corpora, the Bagan
inscriptions, and the Old Javanese Wordnet — the desk opened whole
in one phase), `okhc` — the historical slice of
the Open Korean Historical Corpus, 1,198,779 documents that more
than doubled the library's document count in a day —
`titus-osco-umbrian` (the complete Iguvine Tables with the Oscan
inscriptions, 386 documents at line grain under a personal grant)
and `seal` (Sources of Early Akkadian Literature, 408 compositions,
the editors' written personal-research grant), after the August wave
(the Old Mandarin rhyme books, the Classical↔Modern Chinese parallel
corpus, the Monlam Tibetan lexicon, DACON's Newar, EEBO-TCP,
the Corpus of Middle English, Menota, and the INT Dutch corpora)
and **the granted sources** (2026-07-31): `cantigas`, the
complete secular Galician-Portuguese lyric (1,682 cantigas at verse
grain under the coordinator's written any-use grant) and `rsti`, the
Ras Shamra Tablet Inventory (5,075 inventory cards with KTU/CTA
concordances, published editions at line grain, under Prosser's NC-SA
grant) — the same day Sefaria's classical midrash (Rabbah + Midrash
Halakhah) joined wave 2 and ACTib became nabu-data's first
re-published source. Before them came the P46 wave's `edr`, `ren`, the Geʿez trio, `wold` and
`clics`, after the Romance wave's `openmgh` (the MGH critical editions),
`digiliblt` (late-antique Latin, silver-annotated) and `bfm` (Old
French), joining `glaux` (~20M-token silver-annotated Ancient Greek)
and `croala` (Croatian Latin) from the day before.
The **Arabic library** landed 2026-07-23: `openiti`'s
owner-fired first sync brought 9,079 Islamicate text versions —
34.6M passages of premodern Arabic and Persian. Before it, the
**Germanic four** (`menotec`, `helipad`, `rem`, `rundata`) joined
2026-07-22 with their owner-verified first syncs, a day after `aozora`
(16,004 public-domain works, 2.98M passages) — the upstream corpora
plus four local shelves (the language
dossiers and source dossiers live, the owner's library shelf holding its
first 20 ingested documents, the notes shelf awaiting its first
annotation). The etymology desk now hears
from three tiers of witnesses: the Wiktionary-derived chains, the
expert-curated trio synced 2026-07-14 — **IE-CoR** (4,981 cognate sets /
26,325 reflex rows, 2,308 loan-flagged), **LIV-LOD** (305 PIE verbal
etymons), **de Vaan's Etymological Dictionary of Latin** (2,860
etymons) — and, since 2026-07-17, the five **StarLing / Tower of Babel
bases** under a written grant: Pokorny's complete IEW (2,222 roots),
Nikolayev's PIE database (3,291 etymologies), **Vasmer's etymological
dictionary of Russian** (18,239 entries), and the Common Germanic and
Baltic databases. The same sync wave landed the **Slovenian historical
dictionary shelf** (Pleteršnik 1894–95, the Svetokriški Baroque lexicon,
and the complete 16th-century word inventory — 139,405 entries), the
**damaskini** Balkan Slavic corpus, and the library's **Celtic axis**:
CorPH's gold-lemmatized Early Irish, the RIIG Gaulish inscriptions, the
Ogham in 3D stones, three attested-Celtic Wiktionary extracts, and two
Old Irish UD treebanks.
A local **language-dossier shelf** holds the library's language curation
as 199 plain-Markdown files (`local/shelves/local-language/` — edit in any
editor, `nabu sync local-language` re-derives the `nabu language` cards),
and a **source-dossier shelf** carries a curated description of every
registered source, served on `nabu list` and gate-checked against the
shelf map by `rake site:check`.
Ranked expansion candidates live in the per-axis surveys (Old English,
Slavic, PIE, …) — private planning material under gitignored
`.docs/surveys/`; their license-checked verdicts land in
[docs/02-sources.md](docs/02-sources.md).
There is also a shelf for your own PDFs, scans and articles
(`local/shelves/local-library/` — manifest-catalogued, `research_private` by
default so nothing you scanned is ever served or redistributed,
page-cited where a text layer exists). `nabu ingest FILE...` is its front
door — and it takes http(s) URLs too, downloading first (redirects
followed) and recording the URL you gave in the manifest's `source_url:`
lane: it copies the file in (never moves your original), derives metadata
candidates mechanically (PDF metadata, filename, sha256), walks you
through confirming them — interactively, AI-assisted (`--assist
script/ingest-assist-claude` prefills the prompts with a model's
suggestion), or scripted (`--yes` plus flags) — then syncs the shelf and
prints the minted urn:
```
bin/nabu ingest ~/scans/vaillant-1950-manuel.pdf --collection slavistics
```
After that the manual is `show`-able by page, searchable where it carries
text, and `links`-wired to the passages its manifest entry names as
`related:`.
## Feature tour
| | |
|---|---|
| `nabu search QUERY` | FTS5 full-text search, bm25-ranked, diacritic-insensitive with per-language folding: `μηνιν` finds `μῆνιν`, `iuvenis`/`juvenis`/`iuuenis` all resolve. Filters: `--lang`, `--license`, `--source SLUG` (one shelf) and `--axis NAME[,NAME…]` (a research desk's shelves — the multi-source generalization of `--source`, e.g. `--axis celtic`), both composing with every other filter, `--lemma`/`--near`/`--fuzzy` included, the footer naming the desk; `--limit`. Date/place axis (163,821 dated/placed documents live — EDH inscriptions 81,416, HGV papyri, ORACC catalogue/regnal dates 21,558, TOROT chronicle annals, Slovene goo300k/IMP, Coptic manuscript dates — so `--century -7` reaches the Assyrian letters): `--from -300 --to -30` scopes by signed historical year (negative = BCE, no year 0), `--century 6` is one century's shorthand, `--place oxyrhynch%` filters provenance — `στρατηγ* --from 101 --to 300 --place oxyrhynch%` finds the Oxyrhynchite strategoi. Genre facets (256,518 rows live from EDH): `--type epitaph --province Britannia --material marble` compose with a text query and all of the above. `--kind CLASS[/SUB]` filters by the ruled cross-corpus document-kind classes (1.7M classified documents): `--kind funerary` finds epitaphs across EDH, EDR, IIP and the rest regardless of each collection's own vocabulary (a head matches its whole family; `--kind legal/contract` narrows; `--kind unmapped`/`unknown` are first-class). Character-structure filters for the Han corpus (kept distinct from text FTS): `--radical N` (KangXi radical, Unihan), `--strokes A-B` (total-stroke range), `--char-component C` (characters containing C — KRADFILE ∪ BabelStone IDS transitive containment) AND together and compose with a text query as character-level filters on Han passages, the footer naming them. `--words` (Tibetan): keep only hits whose matched span aligns with WORD boundaries per the published segmentation model — an accidental syllable run across a word boundary no longer matches. |
| `nabu search --lemma FORM` | Dictionary-form search over 21.9M gold lemma rows in 45 languages (2026-08-31 census) — inflections, suppletion and all; hits carry glosses where the reference shelf knows the lemma. Add `--morph case=dat,number=pl` (UD feature vocabulary) to keep only attestations with that morphology, decoded evidence shown per hit — one façade over UD `feats` and PROIEL positional tags. Beside gold and the labeled silver shelves, the **equivalence** tier: CEIPoM's scholar-curated Classical-Latin keys on pre-Roman Italy's Oscan/Umbrian/Faliscan passages — `search --lemma precor` reaches the Iguvine Tables' `pesnimu`, every hit tagged `[equivalence]`, never counted as attestation, `--gold-only` excludes. |
| `nabu search --loans CODE` | The language-contact facet (Coptic Scriptorium today): keep only passages carrying loanwords from a donor language — `search ⲛⲟⲩⲧⲉ --lang cop --loans grc` (131K+ Greek loan tokens tagged), composing with text, `--lemma`, `--fuzzy`, `--near` and every catalog filter. `nabu list coptic-scriptorium --loans` prints the donor-language census; `--loans grc` enumerates the most loan-saturated documents. |
| `nabu search A --near B [--window N]` | Proximity search: keep only hits where `B` is within `N` words of `A` in the same passage (FTS5 NEAR over the folded forms; default 10, `0` = adjacent, order-independent). `λόγος --near θεός` is John 1:1; composes with `--lemma` (the anchor expands to the lemma's attested surface forms first: `--lemma λέγω --near κύριος` finds `τάδε λέγει κύριος`) and `--lang`/`--license`/`--limit`. Both terms bracketed in the snippet. |
| `nabu search --fuzzy FRAGMENT` | Damaged-text fragment search: substring matching ANYWHERE in a passage, mid-word included — `']μηνιν αει['` works typed straight off the edition (editorial brackets stripped, then the same per-language folding as plain search). Character-trigram index over the DOCUMENTARY shelves only (papyri-ddbdp + oracc + edh, `fuzzy_index: true` in the registry — corpus-wide would cost 15×), candidates verified by real substring match; every render names the indexed scope. The production index is LIVE (1,713,135 passages indexed as of 2026-07-14, EDH aboard). Fragments need ≥3 characters; composes with `--lang`/`--license`/`--limit`/date-place filters; `--long` prints the whole folded passage. For literary half-memories use plain search or `parallels`. |
| `nabu show URN` | A passage, a whole document, or a citation range (`urn:…:1.1-1.10`) with license, revision, and full provenance trail. `--parallel` pairs the aligned English translation; `--random` pulls something off the shelf; `--segmented` renders Tibetan text word-split (tsheg delimits syllables, not words — the boundaries come from Nabu's own published segmentation model, `nabu sync nabu-data`). |
| `nabu parallels URN` | Passage-anchored intertext: point at one passage and find where the corpus quotes or echoes it — reception discovery, not translation alignment. Query-time over the FTS index (no new schema): the anchor's 4-word grams are phrase-probed, candidates ranked by shared-gram count weighted by rarity, elision folded across editions (so Matthew 4:4 finds LXX Deuteronomy 8:3). One hit per document (duplicate witnesses grouped, loci counted), the shared phrase shown as evidence; a gold-lemmatized anchor also gets rare-lemma "echoes" (re-inflected allusion). `--long` expands the truncated evidence; `--lang`/`--license`/`--limit` scope. `--batch SCOPE` flips the engine to corpus-wide mining: every anchor of a source slug or urn prefix, hits persisted as `kind=parallel` edges in the links journal (top `--per-anchor` per anchor at `--min-score`+, both named in the summary — no silent caps); reruns supersede, interactive output never persists. |
| `nabu formulas SCOPE` | The oral-formulaic reader's mirror of `parallels`, pointed inward: mine the repeated formulas WITHIN a corpus slice (a source slug or a work/urn prefix) — the same gram machinery, counting instead of probing. Homer's `ὣς ἔφαθ' οἵ δ'` (72×), `τὸν δ' ἀπαμειβόμενος προσέφη πολύμητις Ὀδυσσεύς` (50×); the Old English `saga hwæt ic hatte` riddle refrain, `Beowulf maþelode bearn Ecgþeowes`, `awa to feore`. Ranked by count × length — no stoplist, the ranking is self-filtering (a genuine formula out-recurs any function-word run; measured). `--lang` mines one tradition where a source mixes translations; `--gram-size`, `--min-count`, `--limit`, `--long` (every locus). Zero schema, ~0.2 s per slice. `--batch SCOPE` persists the sweep as `kind=formula` edges: a STAR per formula (hub = its first locus in urn order, one edge to every other locus, the gram riding each edge's detail, the count its score — all-pairs would explode quadratically), top `--max-formulas` by rank; reruns supersede, interactive output never persists. |
| `nabu links URN` | The mined cross-reference graph, read back: every batch-produced edge touching a urn, both directions, grouped by kind (parallel, formula, cognate), counterparts resolved to title/language, each kind's evidence rendered natively (a parallel's score, a formula's `“gram” ×count` pointing at the refrain's hub, a cognate's `ref · root [shelf]` meet), provenance footer citing the producer run (scope, params, code version, date). Edges are urn-keyed in their own journal (`db/links.sqlite3`) so they survive `nabu rebuild` untouched; `show` grows a one-line `linked: N formula, M parallel` footer counting the kinds present when edges exist. `--long` lists every edge per kind. |
| `nabu align REF` | One citation across every witness of a registered work (`config/alignments.yml`) — the parallel NT and the Septuagint ↔ Vulgate OT ship as flagships. A whole-chapter or verse-range query clips at 200 refs by default; `--long` lifts that ceiling and renders every ref. `--collate` diffs the witnesses into a compact apparatus (base reading + per-witness divergences) per (language, script) group, with cross-script witnesses rendered undiffed and labelled honestly; `--base LABEL` picks the base. |
| `nabu define LEMMA` | Dictionary-shelf lookup — LSJ, Lewis & Short, Bosworth-Toller, Monier-Williams (193,890 entries live: SLP1 transcoded to IAST, so `define amsa` reaches aṃśa/aṃsa, RV./BhP. citations resolving into the GRETIL shelf at verse grain, MW's own Gk./Lat./Goth. cognate notes feeding `etym`), Wiktionary OCS — entry citations resolved to in-catalog passages/documents. The TLS Chinese shelf resolves at SENSE grain: `define 棄` lists where each sense is attested in the classics (189K sense-level attributions), the Kanripo-held texts resolving to page or document urns and coverage growing as the kanseki waves sync. A leading `*` scopes to the seven reconstruction shelves (Proto-Slavic/PIE/Proto-Germanic plus the four intermediates), whose entries list their descendant reflexes; `--long` expands the truncated "not attested here" list in full, grouped by language. Query expansion reads Nabu's own published datasets back: a Sanskrit surface form reaches its DCS gold lemmas, and a Tibetan verb stem reaches its paradigm lemma (ཕྱིན or `phyin` finds འགྲོ — suppletion only the published verb table knows). |
| `nabu etym LEMMA` | The comparativist's walk: an attested lemma (богъ, guþ) → every reconstruction whose Wiktionary descendants name it → one hop up the proto-to-proto chain, each with cognates and corpus attestation counts. Multi-hop closure (PIE \*per- → PBS → \*pьrstъ → chu/orv in one walk) and per-edge "(loan)" flags are live, and so — since their 2026-07-14 first syncs — are the three independent expert-curated witnesses: IE-CoR's cognate sets ride the same walk (`etym срьдьцє` → \*k̑erd- with grc καρδία ~ lat cor ~ got hairto beside kaikki's \*ḱérd-, curated loan events labelling their edges `(loan)`), LIV supplies the PIE verbal roots, and de Vaan's EDL the lat → Proto-Italic → PIE Leiden chains. `--long` expands every truncated cognate list, grouped by language with each code named inline (`[gkm · Medieval Greek]`; compact is the default). |
| `nabu language CODE` | The code desk reference: any language code the library surfaces — corpus tags (`chu`, `san-Latn`) and the 803 Wiktionary etymology codes in `etym`'s cognate lists (`gkm`, `zle-ort`, `zlw-opl`) — explained on one card: name (from the derived names census, filled at the owner's 2026-07-14 wiktionary resyncs), family, curated historical context (file-backed since P19-1: one Markdown dossier per code under `local/shelves/local-language/`, 199 live — edit in any editor, `nabu sync local-language` re-derives the card), and live holdings (documents/passages, gold-lemma rows, dictionary shelves, etymology edges; zero fields suppressed). An unknown code misses honestly with a family hint. `--list` prints the held languages plus the nabu-lects registry extension (every anchor's minted stage/variety ids — `grc:koi`, `ave:old` — the vocabulary `--lect` and the postures accept); `--long` adds per-source counts and the upstream-code edge split. |
| `nabu info QUERY` | The dossier front door: one command answering "what is this?" across the three desk namespaces, dispatching to the existing cards byte-identically — a single character or Gardiner-style sign code → the `char` card; a known language code or lect id → the `language` card; a script tag or name (`glag` / `Glagolitic`) → the curated script dossier (the same hand-written context published as nabu-data's `mul/script-dossiers`), with the registry's scripts-table row beneath. Earlier rungs win, shadowed readings are footnoted, and anything else misses honestly naming all three namespaces. |
| `nabu axis [NAME]` | The research-axis desk reference (config/axes.yml) in the `nabu language` mold: bare, every desk with its persona one-liner; `nabu axis celtic`, one desk's full card — persona verbatim, membership rationale, every member source with its enablement and live holdings (documents/passages, entries, languages, license mix; a member holding nothing says so), the aggregate gold-lemma coverage across the desk's held languages, and the shipped affordances (`list --axis NAME`, `sync NAME`). An unknown axis misses naming the known set. |
| `nabu char CHAR` | The character desk card: one Han character composed from every shelf the library holds — the join no dictionary site offers. Header (glyph · total strokes · KangXi radical number+name, from Unihan) → IDS decomposition with each component's follow-up commands (BabelStone IDS) → the component index (KRADFILE) → variants (trad/simp/z-variant, Unihan) → readings (kun/on + meanings from KANJIDIC2; Mandarin/Korean/Vietnamese from Unihan) → pedagogy (Jōyō grade · JLPT · frequency) → desk-reference codes (Unicode always; four-corner/SKIP/JIS/dic where KANJIDIC2 holds them) → **the diachronic column no online dictionary carries** (Baxter-Sagart Old Chinese · Qieyun Middle Chinese · the Heian hanzi dictionaries · TLS senses with classical attestation counts) → corpus attestation → the `search CHAR*` follow-ups. Matches Jisho field-for-field where a shelf backs it, exceeds it diachronically, and **never renders an unbacked field** (absent, not "—"). |
| `nabu signs URN\|TEXT` | The cuneiform sign desk: ATF transliteration — a passage/document urn from the ATF shelves (cdli/oracc/ebl/etcsl) or raw pasted text — resolved token by token through the local Oracc Sign List (`nabu sync osl`) into sign identities: value → sign name → Unicode codepoint(s)/glyph. C-ATF ASCII folds (sz→š, s,→ṣ, u2→u₂) and the ETCSL romanization (c→š, j→ŋ; auto for etcsl urns, `--dialect etcsl` for raw text). One honest status per token — deterministic · qualified · ambiguous (**all** candidates listed) · no-codepoint (sign real, honestly unencoded) · broken · unknown (not in the OSL, said plainly); determinatives unbraced and marked; absent data absent, never "—". `--json` emits the frozen machine contract the MCP `nabu_signs` tool serves identically — the downstream boundary: **the sign list adds identity, not curriculum**. |
| `nabu cognates TARGET` | Cognates in parallel: verses of an alignment work where witnesses in ≥2 languages use reflexes of the same reconstruction root — the hub × crosswalk join (`cognates nt --langs got,chu` → salt~соль at \*sḗh₂l, hlaifs~хлѣбъ at \*hlaibaz, ~300 verses / 30 roots in under a second). TARGET is a work id, verse, chapter, or book; every hit names its meet SHELF (a gem-pro meet for a Slavic word reads as a borrowing); corpus-common words suppressed with an honest count (`--all` lifts); `--long` lifts the 200-hit cap and expands detail. `--batch WORK` persists the whole-work map as `kind=cognate` edges between the cross-language witness passages of each meet, the meet itself (`ref · root [shelf]`) riding each edge's detail; reruns supersede, interactive output never persists. |
| `nabu concord QUERY` | Classic KWIC concordance: keyword column-aligned in pristine text, corpus order — for scanning usage, not relevance. |
| `nabu kind census` | The document-kind axis (the fourth axis beside dates, places, and lects): what KIND of document is each text — funerary, administrative, legal, letter, divination, historiography… Thirteen collections' own genre vocabularies (sepulcralis / epitaph / funerary / Inscription funéraire) fold onto ~26 ruled cross-corpus classes, multi-label (an ode to a ruler is poetry AND royal), every upstream label preserved verbatim. `census` shows the class board with document/source spread, the honesty buckets (`unknown` = upstream's own "cannot determine"; `unmapped` = values awaiting a fold rule — `--unmapped` lists them per source), and the unclassified remainder. Filter with `search --kind`; a document's classes render on its `show` card. |
| `nabu vocab URN` | Lemma-frequency profile of a document, range, or passage against the gold-lemma corpus: total tokens, distinct lemmas, the most distinctive vocabulary (log-odds vs corpus — Caesar surfaces *legio*/*proelium*, Cicero's *De officiis* surfaces *officium*/*honestas*), and the in-document hapax legomena. Gold shelves only; a document without gold lemmas says so and names the annotated languages. `--long` lists every hapax (and every gold-bearing language) in full, escaping the `--limit` display cap. `--by-century` switches to diachronic mode: the shape of the dated corpus over time, or — with a text query — a word plotted across the centuries (`vocab --by-century 'στρατηγ*' --lang grc` peaks in the 2nd c. CE), bucketed by earliest year and honest about ranges that span more than one. |
| `nabu lemma FORM [--lang CODE]` | A lemma's attestation profile from the precompiled lemma rollup: first and last attested century, each with its date basis (artifact vs composition dating), a per-century table and sparkline with gold (verified annotation) and silver (automatic lemmatization) kept apart, undated attestations listed last, the top sources, and the dictionary shelves carrying the form. A document counts at the midpoint of its date envelope. Companions on the language desk: `nabu language CODE --curve` (documents + passages per century) and `nabu language --mixing` (the multilingualism census — documents by passage-language inventory size, the commonest mixed inventories, and per source whether mixing runs in blocks or interleaves). All three read only precompiled rollups, rebuilt with the library and by `rake builders:refresh`. |
| `nabu export --format plain\|jsonl` | Stream the corpus out, with `--lang`/`--license`/`--source`/`--axis NAME[,NAME…]` filters (`--axis` is the multi-source generalization of `--source` — a whole research desk's shelves) — the longevity-hedge exit formats. |
| `nabu ingest FILE-or-URL...` | The intake front door for your own material: copies a PDF/scan/article into the local-library shelf (`research_private` by default — never served or redistributed) — or downloads an http(s) URL first, recording it in the manifest — derives metadata candidates mechanically, confirms them interactively / AI-assisted (`--assist CMD`) / scripted (`--yes`), then syncs and prints the minted urn. `--shelf language CODE` scaffolds a language dossier, `--shelf source SLUG` a source dossier. |
| `nabu note URN [TEXT]` | Owner annotations — scholia of one's own — on any urn the corpus knows (documents, passages, ranges, dictionary entries), resolution-checked before any write, stored as plain YAML on the local-notes shelf. Bare `nabu note URN` reads back what you said; `--list` enumerates; `--force` records a deliberately dangling note on planned material. Notes render on `show`/`define`/`links` and are served over MCP with their target's withholding rules. |
| `nabu enable AXIS\|SOURCE...` | The first-time step before `sync`: adds sources and/or whole desks to `local/config/profile.yml`, this box's enablement config — enablement governs what `sync`/`sync --all` acquires and the default row set of `list`/`status`/`health` (query commands always see everything held). Grant-gated sources prompt for the recorded acknowledgment (`--grant-acknowledged` for scripted use). Browse the menu first with `status --disabled` / `list --disabled` — the complement view: only the rows `enable` would add (`--all` is everything). |
| `nabu sync SLUG` / `sync --all` | Fetch and load a source (git, zip, or single-file HTTP — or re-scan a local shelf); idempotent, non-destructive, every run recorded. The name resolves slug-first-then-axis: `sync celtic` (or `--axis a,b`) expands a research axis to its enabled members, grouped, and names any unwired members on one skip line — feature modules on their own `skipped (module — sync directly)` line — (an axis is not an explicit request, so an unwired member is skipped where an explicit `sync SLUG` would sync it anyway). |
| `nabu list [SOURCE]` | The what-is-held view (`status` is the sync-state view): bare, a content census — one line per shelf with document/passage/entry counts, languages, the effective license-class mix, withdrawn/retired counts when nonzero. With a SOURCE, one shelf's card (identity, credit line, counts, per-language breakdown, dictionaries, date-axis coverage, facet and collection summaries). `--documents` / `--entries` / `--collections` enumerate (default `--limit 50`, `0` = all, honest "… N more" tail), with `--lang`/`--license`/`--withdrawn`/`--from`/`--to`/`--century` filters on documents and `--prefix` folded headword-prefix filtering on entries (`bh` finds `*bʰer-`). `--axis` groups the census under the research axes (the owner's desks) — bare, every axis in ratified order with its persona line; `--axis slavic` or `--axis a,b` selects some — a source appearing under each axis it serves. |
| `nabu status` / `health` / `verify` | Per-source counts and run history, each row carrying an `up=` upstream-drift column (`up=ok(2d)` / `up=BEHIND(2d)` / `up=stale(30d)` / `up=?(never)` / `up=?(re-probe)` when a cached verdict predates the last sync / `up=frozen`) so an update is an informed decision — `nabu status --remote` probes upstreams inline and refreshes it in one command; local trend + upstream drift checks; full bitrot/tamper re-verification of every canonical file. `health` also runs the mechanical postcondition invariants (P18-7): failed-run/partial-load surfacing, flag-vs-artifact and synced-vs-populated mismatches, pending migrations, and quarantine counts as a DELTA against an audited baseline — plus an optional `sync --review CMD` AI-review hook, off by default. Its default view prints findings only (ok rows and by-design notes fold into one rollup; `nabu health --all` shows every row). |
| `nabu mcp` | The read-only MCP server — fourteen tools for Claude Code/Desktop and any MCP client. Recipes and the user-perspective walkthroughs in [docs/mcp.md](docs/mcp.md). |
Two more tastes. Facing translation, span-grouped, honest when the English
is coarser than the Greek:
```
$ bin/nabu show urn:cts:greekLit:tlg0012.tlg001.perseus-grc2:1.1-1.5 --parallel
urn:cts:greekLit:tlg0012.tlg001.perseus-grc2 — Iliad [grc]
parallel: urn:cts:greekLit:tlg0012.tlg001.perseus-eng4 — Iliad [eng]
aligned by citation: 0 paired, 1 block covering 5 lines, 0 grc only, 0 eng only
:1.1
grc μῆνιν ἄειδε θεὰ Πηληϊάδεω Ἀχιλῆος
:1.2
grc οὐλομένην, ἣ μυρίʼ Ἀχαιοῖς ἄλγεʼ ἔθηκε,
…
eng [:1.1 — covers :1.1–:1.39; range shows :1.1–:1.5]
Sing, O goddess, the anger [mênis] of Achilles son of Peleus, that brought
countless ills upon the Achaeans. …
```
And the concordance, here on Caesar's *virtus*:
```
$ bin/nabu concord --lemma virtus --width 30
…vetii quoque reliquos Gallos virtute praecedunt, quod fere cotidi… urn:nabu:proiel:caes-gal:52552 [lat]
…i populi Romani et pristinae virtutis Helvetiorum. urn:nabu:proiel:caes-gal:52635 [lat]
…b eam rem aut suae magnopere virtuti tribueret aut ipsos despicer… urn:nabu:proiel:caes-gal:52636 [lat]
…que suis didicisse, ut magis virtute contenderent quam dolo aut i… urn:nabu:proiel:caes-gal:52637 [lat]
…
```
## Your collection cannot rot
Upstream projects restructure, lose funding, and disappear. Nabu is built
on the assumption that the library must outlive its sources:
- **Canonical vs. derived.** Upstream text lives as plain files in a
git-tracked canonical layer — the permanent asset. All SQLite is derived
and rebuildable: `nabu rebuild` regenerates the entire catalog from
canonical data, proven byte-identical by test.
- **The attic.** Fetch is non-destructive: files an upstream deletes are
copied to `canonical/<source>/.attic/` *before* the merge and stay live,
searchable, and exportable — honestly labeled "retired upstream", keeping
the license they were fetched under. A mass-deletion breaker aborts any
sync that would withdraw more than 20% of a source.
- **The ledger.** Run history, license baselines, and revision records live
in `local/history.sqlite3` — the instance ledger no rebuild touches,
preserved by backup as part of `local/` (the P71 three-folder contract).
- **Backup with a drill.** `nabu backup` rsyncs the three permanent
folders (`canonical/` + `config/` + `local/`) — plus the derived dbs as
a convenience — to a mounted external volume; `rake ops:drill` proves
it (backup → fresh-root restore → rebuild → verify → RESTORABLE), and
`rake ops:drill_pure` proves the derivability law itself: restore the
three folders ONLY and re-derive every database from scratch.
- **Standing verification.** `nabu verify` re-parses every canonical file
(attic included) and compares content hashes against the catalog;
`nabu health` watches run-history trends and probes upstreams for drift.
Every sync prints a discovery-accounting line (`selected ·
skipped-by-rule · unrecognized`), so silent ingestion gaps are
structurally visible.
- **Boring storage.** Files, git, SQLite. Restorable from an rsync with
zero services.
Nothing is ever hard-deleted: withdraw, revise, journal.
## Status — honest
This is a **young, personal, early-development project** — built for one
scholar's research needs first and shared because the approach may be
useful to others.
- Developed and tested on macOS (Apple Silicon), Ruby 3.3+. Nothing is
known to be Mac-specific except the ops templates (launchd), but no other
platform is exercised.
- Versioned releases begin at **v1.0.0** (cut at the Phase-19 gate;
citation metadata ships in [CITATION.cff](CITATION.cff)); there is no
gem and CLI flags may still change between releases. GitHub Actions CI
runs the full suite plus rubocop (`rake test` + `rake lint`,
network-blocked, fast) on every push and pull request — the badge up
top is the contract.
- Corpus numbers above are a snapshot of one live install, dated where they
appear.
- The enrichment layer of the original vision (embeddings/semantic search,
machine glossing) is designed but not built — see
[docs/01-concept.md](docs/01-concept.md) for where this is headed.
Ad-hoc ingestion of your own material shipped as `nabu ingest`
(Phase 19); OCR/HTR for image-only scans still waits on local inference
hardware.
- Expect rough edges; expect the docs to be more honest than polished.
## The project family
Nabu has three public siblings:
- **[nabu-data](https://github.com/arvicco/nabu-data)** — the datasets
Nabu itself publishes: form→lemma tables, orthography and script folds,
metrical scansions, the Tibetan segmentation layer, the cuneiform
value→sign table, the curated language and script dossiers, the corpus
layers (stages, dates, places, signs), and the re-publications (the
ACTib anchor layer, the complete secular Galician-Portuguese lyric) —
twenty-three datasets as
plain CSV with Frictionless Data Package manifests (CC BY 4.0, stated
share-alike carve-outs), each carrying its full derivation provenance
down to the input corpus shas, archived with a version DOI under
concept [10.5281/zenodo.21757475](https://doi.org/10.5281/zenodo.21757475).
The loop runs both ways: Nabu re-ingests its own publication as an
ordinary source (`nabu sync nabu-data`), so the Sanskrit lookup lane,
the Tibetan verb lane, and Tibetan word segmentation are all served
from the published files.
- **[nabu-lects](https://arvicco.github.io/nabu-lects)** — the lect
registry ([repository](https://github.com/arvicco/nabu-lects), CC BY
4.0, pre-1.0): language varieties as genealogical anchor × historical
stage × variety × orthography, with a universal mapping from standard
codes onto them. Born from this library's holdings, universal by
construction; Nabu consumes it (`nabu sync nabu-lects`) for the stage
ladders on `nabu language` cards, `search --lect`, and registry-keyed
reconstruction display (docs/languages.md, "The lect layer").
- **[Edubba](https://arvicco.github.io/nabu-edubba)** — the scribal
school: a downstream teaching project built on Nabu's library, named
for the Sumerian é-dub-ba-a, the "tablet house". Where Nabu curates
the texts, Edubba teaches the scripts; its reading panels consume
Nabu's MCP and sign-list surfaces. The boundary is deliberate: the
library adds identity, not curriculum.
## Documentation
| Doc | One line |
|---|---|
| [docs/quickstart.md](docs/quickstart.md) | Zero to first search, copy-pasteable, honest about sizes and timings. |
| [docs/library.md](docs/library.md) | The shelf map: every corpus with contents, counts, licenses, and research uses. |
| [docs/axes.md](docs/axes.md) | the research axes: the twenty-four scholarly desks, their personas, and which sources each tags. |
| [docs/01-concept.md](docs/01-concept.md) | The vision: what Nabu is, workflows, principles, what success looks like. |
| [docs/mcp.md](docs/mcp.md) | The MCP server: twelve read-only tools, registration recipes, quoting etiquette. |
| [docs/conventions.md](docs/conventions.md) | Field notes for working with ancient-text corpora (Unicode/NFC, citations, editions, licensing) — start here if you're new to the domain. |
| [docs/architecture.md](docs/architecture.md) | The design: layer model, adapter contract, store schema, retention machinery. |
| [docs/02-sources.md](docs/02-sources.md) | The source inventory: every corpus scouted, scored, and license-checked. |
| [docs/03-unlockable-sources.md](docs/03-unlockable-sources.md) | Sources not ingestible today, with concrete unlock paths. |
| [docs/ops.md](docs/ops.md) | The runbook: maintenance cadence, launchd templates, what to do when a check goes red. |
| [docs/maintenance-and-extension.md](docs/maintenance-and-extension.md) | How this stays alive across years of intermittent attention — leads with the **where-truth-lives map**: which surface is authoritative for every aspect, and what merely projects it. |
## How this is built
Nabu is developed by a model-tiered autonomous agent loop — work packets
executed by Claude models under TDD ground rules, with owner-approved phase
gates — documented in [docs/dev-loop.md](docs/dev-loop.md). This README is
refreshed at every gate to reflect what actually works.
```
bundle exec rake test # full suite (network-blocked by WebMock; fast)
bundle exec rake lint # rubocop
bin/nabu --help
```
Contributions: the project is early and personal; issues and conversation
are welcome, but expect the backlog to be driven by the owner's research
needs. The house rules for outside contributors — TDD, fixture discipline,
the DCO sign-off — and the issue templates for requesting sources and
features or reporting a wrong reading are in
[CONTRIBUTING.md](CONTRIBUTING.md); if you want to add a source,
`CLAUDE.md` and
[docs/maintenance-and-extension.md](docs/maintenance-and-extension.md)
describe the adapter checklist end to end.
## License
- **Code:** [MIT](LICENSE).
- **Content:** every ingested text keeps its upstream license, recorded
per document as data (`open` / `attribution` / `nc`), and every surface —
search hits, exports, MCP responses — carries the label. Roughly 99% of
documents are public-domain or attribution-class; the `nc` shelves
(GRETIL, most treebanks) are for non-commercial research use and are
never redistributed by the tooling. Per-source terms:
[docs/02-sources.md](docs/02-sources.md).