A digital
encyclopedia

Every entry on Alexandria sits in one of four tiers. The base — Tier 0 — is a stub: a name, a span of dates, a handful of relationships pulled from Wikidata. Stubs surface only through other entries; they never appear on the homepage or in search results above the fold. They mark the spaces the Library is still growing into.

Tier 1 entries have a 150–300 word summary, adapted from Wikipedia's lead section through a rewrite layer. Tier 2 entries have a full 800–1,500 word narrative, synthesised across Wikipedia and (where available) the 1911 Encyclopædia Britannica — multi-source synthesis rather than paraphrase. Each Tier 2 piece is fact-checked by a separate model run; flagged claims are surfaced on the entry itself.

Tier 3 is the curated layer — hand-picked imagery, longer-form prose (2,500–3,500 words), no algorithmic caps. There are five Tier 3 entries at launch; more will arrive as the editorial work continues.

Most encyclopedias of human history default to the West. Wikipedia is denser in English; English is densest on European subjects; the long tail of non-European history is underrepresented even when the underlying scholarship is rich. Alexandria explicitly resists that default at every layer: the seed filter rewards entries with sitelinks in non-European Wikipedias, the homepage rotation never lets one civilisation dominate, civilisational tags are curated rather than imposed from the UN subregion map, and the calibration set the prose was tuned against runs from Hannibal to Mansa Musa to Wu Zetian to Tupac Amaru II.

Beyond that: the prose is meant to read like the best long-form journalism, not like a textbook. Each entry has permission to lead with a paradox or an image, develop context at length, and close on a question that opens to other entries. Where the sources disagree, the disagreement is named rather than papered over.

A continuously-running pipeline upgrades entries through the tiers as inbound interest, source density, and editorial attention warrant. Summarisation and narration are done by Google's Gemini 3.5 Flash via the Vercel AI Gateway, with a hard daily budget cap. Embeddings (used for semantic search and the “Adjacent in spirit” widget on every entry) are Voyage 3 large, 1,024 dimensions, cosine-ranked through pgvector with HNSW indexing.

The prose layer is constrained: cite only facts present in the provided sources; never invent dates, quotes, places, borders, or scholarly attributions; flag disagreements briefly rather than picking a side and burying the other. The fact-check pass that runs after each narrative is a second model run reading the same sources, asked to surface every claim the sources don't support. Findings get published alongside the narrative — readers see what the system itself is uncertain about.

Wikipedia content is adapted under CC BY-SA 4.0. The 1911 Encyclopædia Britannica is public domain in the United States; entries are pulled via Wikisource. Imagery is from Wikimedia Commons, with per-image attribution and license shown on each entry.

The code is open source at github.com/chloeilabs/alexandria. The stack is Next.js 16 + Postgres 17 with the pgvector extension + Drizzle ORM, deployed on Vercel with Neon as the production database.

Try the threads — curated paths through five to seven entries that tell a continuous story. Or browse civilizations if you want to read across one part of the world. Random drops you on a curated entry; from there, the connections link onward.