Every exchange's narrative in this database is vectorized (TF-IDF) and clustered by vocabulary, then plotted in time and space. Click any point to see its nearest stylistic neighbors, or read how this was built in the method note below.
Each exchange's architecture text (style, notable features, and the building narrative) or full narrative (all five research categories) is vectorized with TF-IDF (unigrams and bigrams, English stopwords plus citation boilerplate removed), with each entry's own city, country, and name redacted first so a match reflects shared vocabulary about materials, forms, and typology — not two entries simply mentioning the same place. Similarity is cosine distance between TF-IDF vectors; clusters are k-means (k=8) on the same vectors, labeled by the terms most over-represented in that cluster relative to the whole corpus.
This measures how similarly the prose describes a building, which depends on what sources exist and how thoroughly each entry has been researched — not a direct measurement of the physical architecture. Treat clusters as a lead for further art-historical comparison, not a conclusion.