Data sources

Processed Data Sources

In the kunst.bild.daten project, several curated ZI data holdings and open authority datasets are brought together to consolidate the heterogeneous data landscape of the Photothek, which has grown over decades.

ZI’s Own Holdings

  • ZI Photothek — The core holding is the handwritten and typewritten inventory from the years 1947–1999: approx. 66,000 pages with approx. 529,000 cards documenting around 700,000 inventory numbers, supplemented by 5 handwritten list volumes from the early years
  • HIDA-Midas Database (2000–2015) — earlier digital capture
  • Google Arts & Culture Brief Records — Excel spreadsheets from the Google digitization project (2016–2025), data on the topography of Germany (TOPD) as well as on artists (KUE)
  • Inventory Lists for Smaller Sub-Holdings — series-specific holdings lists (KGW, LEHM, MUSTOP)

External Authority Data

  • Wikidata — filtered dump (latest-all.json.bz2, ~90 million entities), from which an art-relevant subset is pre-filtered using over 69 curated properties. For artist authority-data matching, a person subset of around 980,000 entries was generated from this via external identifiers (including AGORHA ID, Artists of the World ID, RKDartists ID). Wikidata serves here as an alignment layer, not as a reference work, and is merged with ULAN and Foto Marburg into a consolidated resource (1,187,438 documents). For buildings, however, Wikidata is not used — the name forms could barely be reconciled with the ZI data; instead, OpenStreetMap and the Wikipedia index are used here (see below)

  • Getty ULAN — Union List of Artist Names; approx. 176,000 entries (133,006 exclusively via ULAN, the rest in combination with Wikidata/Foto Marburg) feed into the consolidated artist resource

  • Foto Marburg / Bildarchiv — artist documents (fm_kuenstler-combined.xml), approx. 96,500 entries feed into the consolidated resource

  • OpenStreetMap (Geo Index) — local index for location search and geo-coordinate enrichment of buildings; provides spatial distance as a linking point for the downstream Wikipedia matching

  • German Wikipedia — dewiki_content full-text index (Elasticsearch), for disambiguation and geo-linking of buildings (piloted on the topography of Germany: around 18,000 buildings in approx. 6,000 locations)

Models and Services

  • Google Gemini 2.5 Flash — index card recognition and field extraction (two-stage pipeline: layout-based extraction → conversion into a Pydantic data model with 22 fields)
  • Local Qwen Model — search string generator in the geographic agent system (normalization of building names and generation of name variants for Wikipedia matching)
  • Elasticsearch / OpenSearch — central search and reconciliation indices (facet search, artist reconciliation, Wikipedia and geo index)
  • DuckDB — analytics layer