The search index behind digital-collections-explorer.com: 562,842 images of 56,554 items from the Library of Congress Geography and Map Division, most of them Sanborn fire insurance maps. Download it and run Digital Collections Explorer to search the maps by what is drawn on them.
Then open http://localhost:8000. The first start downloads the CLIP model, about 600 MB, so it takes a few minutes. The site needs about 3 GB of memory. The maps are not in the download: each one is loaded from the Library of Congress's IIIF image servers when it is shown.
embeddings.pt is beto_normalized.pt from clip-loc-maps by Jamie Mahowald (2024), unchanged. item_ids.pt and metadata.json were made from that record's beto_idx.pt and merged_files.csv by create_loc_assets.py. The titles and dates were read from the loc.gov API in November 2026 by fetch_loc_titles.py. 28 items, with 1,682 images, were no longer at loc.gov then, so their images have no title or link.
Huang, Ying-Hsiang, and Benjamin Charles Germain Lee. 2025. "Digital Collections Explorer: An Open-Source, Multimodal Viewer for Searching Digital Collections." Computational Humanities Research 1: e14. https://doi.org/10.1017/chr.2025.10017
Mahowald, Jamie. 2024. clip-loc-maps. Zenodo. https://doi.org/10.5281/zenodo.11538437
CC BY 4.0, the license of the embeddings it is built on.