Eleven volumes of the Main Catalogue were compiled by many hands across more than half a century, each cataloguer describing manuscripts in rich but idiosyncratic scholarly prose. Imposing a fixed schema on that material up front would have forced decades of describing practice into categories it was never written for, quietly discarding whatever did not fit. The motivation for a data-driven approach was the opposite instinct — let the corpus declare its own structure, so the model answers to the manuscripts rather than the manuscripts to the model.

The data model was derived from the material itself, not imposed on it. A custom analysis of the full 4.3-million-word corpus surfaced 141 candidate description sections — content descriptions, colophon types, structural identifiers — and shaped every modelling decision. An agentic-AI extraction layer (LLM agents driven by prompt chains and validation steps) then turns the free-form scholarly prose of eleven volumes into typed, queryable fields — sections, colophons, scribes, dates, materials — normalising redundant and denormalised variants, with domain experts reviewing every entry before it is published.

Letting the data drive the design means the schema fits how cataloguers actually wrote across decades and volumes, so inconsistencies are absorbed rather than lost. The result is precise faceted filtering (by scribe, colophon, century and more), reliable full-text, fuzzy and proximity search, and a clean public API for programmatic research — over a foundation that also unlocks computational methods such as classification, cross-referencing and knowledge-graph exploration that flat catalogue text could never support.

The approach trades an easy start for a faithful result, and the difficulty lives in that trade. Scholarly description is inconsistent by nature — the same feature named differently across volumes, colophons and content descriptions interleaved, dates and materials recorded in prose rather than fields — so much of the work is normalising redundant and denormalised variants without flattening the distinctions that matter. The agentic-AI extraction had to be constrained hard against plausible-but-wrong readings: prompt chains, typed schemas and validation steps catch what free-form generation would otherwise invent, and every entry still passes domain-expert review before publication. The schema itself keeps moving as new description patterns surface in later volumes, so it is versioned and migrated rather than frozen — and all of it must hold across Classical and Modern Armenian orthography, transliteration and diacritics that general-purpose tooling handles poorly.

armcatalogs.matenadaran.am opens the Matenadaran’s Main Catalogue of Armenian manuscripts — 11 printed volumes, 3,700+ manuscripts and 4.3 million words of scholarly description — to structured, worldwide access. Previously available only in print and as flat website text, the catalogue is now a queryable dataset behind full-text, fuzzy and proximity search (Elasticsearch), faceted filtering, and a public REST API.