Three windows since v1.7.0 — the library learns to answer
22 October 2026 · Nabu news
The release cut on 20 October was an engine milestone; the two days since have been about what the engine can now say. Three windows in three days — and a library standing at 2,854,118 documents / 128,499,000 passages across 327 live sources (as of 7 November 2026).
First, the entity layer became readable. The person and place
names that parsers had been quietly indexing (318,542 name–passage
pairs) got their own desk: nabu entity looks a name up across every
tradition at millisecond speed, show --entities reads a passage’s
marked names back, and dictionary cards climb their own etymon
chains. Beside it, the dating season continued — a regnal table for
the Pallava and Chola inscriptions added 451 dated documents, each
accession year citing its published chronology — and the Baltic
littoral opened: the earliest Latvian prints (1585–1685), the Old
Latvian sermon corpus, Basque’s first books staged by era, and an
86,067-entry Classical Armenian lexicon. South Indian epigraphy
widened by 18,495 documents under a repository-wide grant read
verbatim from the maintainers.
Then the content wave. Eighteen new sources and two canon widens in a single window: the VedaWeb Rigveda with gold morphology on every word; the Liège LASLA corpus of gold-lemmatized classical Latin (Plautus to Juvenal); Lotsawa House’s Tibetan Buddhist translations in nine languages; the Anyang oracle-bone corpus — the earliest Chinese stratum, 10,077 rubbings; the Shuowen Jiezi wired into the character card; Sefaria’s medieval library (the Rishonim, Mishneh Torah, the Zohar) staged as Medieval Hebrew; ten historical editions of the Chinese Buddhist canon; seven South-Asian dictionary shelves; five more Wikisource language cones from Telugu to Breton; and the Caucasus instruments — Old Georgian and Middle Armenian treebanks on the registry’s new anchors.
And now the reading instruments. Today’s window opened four doors into what those waves had already fetched. Translations that ride inside annotations — the cuneiform corpus’s translation lines, Coptic Scriptorium, the Hittite glossed corpus’s German and eleven more lanes, about 302,000 passages — answer full-text search and point back at their originals. Six more morphology tagsets decode through the search façade (the Diorisis Greek analyses, the ETCBC Hebrew of the Bible and the Scrolls, OpenScriptures Hebrew, two medieval German corpora, Old Norse manuscripts), with upstream word glosses joining the search beside them. The Digital Corpus of Sanskrit’s 531,573 sense-tagged tokens became the library’s first word-sense search. And the cuneiform character card now carries the printed sign-list concordances — Labat, ATU, MZL and kin — 17,090 rows over 2,555 signs, precompiled.
Underneath, the honesty machinery kept pace: fetches of the scan-indexed Wikisource family are now atomic with their cone membership persisted (a mass-deletion breaker caught the gap, working as designed), the Sefaria kind fold reads the deep category path so commentary classes as commentary, and the test suite — 14,471 tests — runs in ninety seconds where it took seven minutes.