10 May 2026
Source Library, by the Numbers
A snapshot of the corpus: what's in it, where it came from, and what it takes to hold it together.
Source Library holds 26,643 works in total: 11,965 books and 14,412 artworks. The numbers below describe the book corpus unless noted.
3.83 billion words on disk
From scan to readable text
Every book page travels a pipeline: photographed, OCR'd, translated, indexed. Here's how far the 4,294,562 pages of the book corpus have come.
199 languages, from Sumerian to Russian
Books in their original languages, before AI translation. The corpus spans 199 distinct languages and language combinations; the top fourteen are below. Latin and German anchor the European corpus; ETCSL contributes Sumerian; dedicated import streams bring in Chinese, Sanskrit, and Hebrew.
Across nearly five thousand years
Distribution of works by century of original composition. The earliest item dates to 2880 BCE (a stela of King Raneb of Egypt's Second Dynasty), and the corpus runs unbroken through to the present. The book corpus peaks in the 1600s (the heyday of Hermetic and alchemical printing), with strong shoulders in the 1500s and 1800s.
Where the books come from
Source Library mirrors and unifies digitized holdings from libraries and museums around the world. The provenance of every page is preserved.
What it takes to hold it
Three primary stores keep the corpus alive: a metadata database, a search-and-text mirror, and an image archive.
- 5.7M page records · 41 GB
- 82K chapter texts · 6.5 GB
- 811K entities · 1.4 GB
- 92 collections, 25.5M documents total
- Full-text search across every translated page
- 30,564 book-level vector embeddings
- Per-page CLIP image embeddings for visual search
- 7-lane semantic search powers /search
- 3.9M archived full-res JPEGs (~286 KB each)
- 3.9M 1200px display images + thumbnails
- 100K extracted illustrations & artworks
- BPH JP2 originals from Embassy of the Free Mind
Generated by AI
Every translation, summary, and chapter index in Source Library is produced by Google Gemini models. The original text is always preserved, one click away on every page.
A knowledge graph of 811,127 entities
Every translated page is scanned for the people it names, the places it mentions, and the concepts it invokes. The result is a connected graph that links Hermes Trismegistus to Marsilio Ficino to a print on the Rijksmuseum's wall.
Tokens, end to end
Reading the corpus and translating it has cost about 27 billion Gemini tokens across every phase of the pipeline.
Tokens by pipeline phase
How this article was made
This is an AI-assisted article. Derek Lomas set the question; AI models searched the library, read the sources and drafted the text, which was reviewed before it was published.
Every quotation is checked by machine against the scanned page it comes from and links to that page, so you can read the original yourself. The article is a living document: it is revised when a reader finds an error or the library gains a better source.
How these articles are made · Last revised 18 October 2026 · Revision history · Spot an error? Suggest an edit
Related articles
More from the same category first, then the newest articles.
