The library steps outside — a public demo image, and the languages of ancient Italy
30 October 2026 · Nabu news
Four development windows since the last note, and the first one that points outward: the library now has a public, runnable demo.
One command, a working research library. The demo image at
ghcr.io/arvicco/nabu-demo bakes a fully open slice of the library —
223 documents, 104,585 passages, 168,142 dictionary entries, every
source openly licensed and attributed — behind the same MCP server
the full library runs. docker run -i --rm ghcr.io/arvicco/nabu-demo
gives any MCP client (Claude, Cursor, VS Code and friends) fourteen
working tools over real corpus data: search ἀγάπη and land in
1 Corinthians 13 with the concordance, definitions and alignment
witnesses beside it. The server also speaks streamable HTTP for
hosted setups. The listing is live on Glama, with more directories to
follow; the full library remains a local build from the open adapters
in the repo. Details on the MCP page.
Eleven languages of ancient Italy entered the registry. The Lexicon ItAnt — 809 dictionary entries across sixteen languages of pre-Roman Italy, from Oscan to Venetic — joined the dictionary shelves, and the lect registry grew the anchors its languages needed: Faliscan, Venetic, South Picene, Paelignian, Marrucinian, Marsian, Vestinian, Volscian, Hernican, and local-use codes for Pre-Samnite and Capenate, whose tags ISO 639-3 never assigned. Each anchor carries researched classification notes and an epigraphic attestation band; Venetic’s disputed place inside Italic is recorded as disputed rather than silently decided. The registry now holds 284 anchors.
New on the shelves. Théâtre Classique brought French classical drama — Corneille, Molière, Racine and their contemporaries — as structured play texts. A Literary Chinese collection widened the largest shelf. Jäschke’s 1881 Tibetan–English dictionary arrived as a scan-anchored reference: page-exact lookups from the dictionary desk into the original plates, with a converter verified against the scans themselves. Sefaria’s curated cross-reference graph now rides the Tanakh shelf, linking verses to the passages that cite them. And the reference-scan shelf gained its Celtic wing: Zeuss and Ebel’s Grammatica Celtica (1871) and the held volume of Holder’s Alt-celtischer Sprachschatz.
The person index learned Assyriology. The Prosopography of the Neo-Assyrian Empire joined the person authorities, with its persons linked to the texts that attest them — the index that already serves Greek and Latin epigraphy now answers for Nineveh too.
An experiment, reported honestly. The parallels engine gained a contiguity bonus last window (adjacent shared phrases now outrank scattered ones, measured as a recall gain). This window tested a fourth idea — demoting phrases a document repeats internally, the “framing formula” discount — on all five gold sets. It never helped: on the Pali suttas it was built for, the framing formulas turn out to carry the curated-parallel signal. The switch stays in the bench harness, off, with the measurements in the repo; negative results are results.
The shelves stand at 2,854,118 documents / 128,499,000 passages across 327 live sources (as of 7 November 2026).