Site icon The GDELT Project

Gemini For Museums: Collection-Scale Cataloging Of Handwritten Annotations

Books that have survived centuries of use are often filled with hidden handwritten annotations, ranging from the usual ownership signatures and remarks to missing page numbers to underlined passages to lengthy marginalia and passages. Continuing our new collection-scale analysis workflow built on Gemini Carbon, we've added a new module for collection-scale marginalia and annotation detection, applying it to a 1616 printing of Delrio's Disquisitionum Magicarum. Here Gemini struggles and we tested multiple iterations, trading off recall (finding all annotations) with precision (avoiding as many false positives as possible). The final module's results, seen below, balance these two goals reasonably well, correctly detecting several things we had missed, including the added page numbering of the index, but also mistaking several printing artifacts for annotations. We'll continue to refine this module.

We've added in token consumption reporting to each task, with this task taking 1,541 seconds (~26 minutes) to examine all 1,081 pages, making 73 API calls and consuming 2,125,400 input tokens, 21,705 output tokens and 90,127 thinking tokens. We make full use of the chunk caching abilities of the Interactions API, dramatically reducing token costs when running all of the tasks sequentially.

Exit mobile version