A graph you can look at.
Start from Babylon or another example, laid out as a force graph. Lines are colored by kind of relation. Dashed ones are facts the graph's own type checks flagged as out of range: shown, not hidden.
In the sample
Click an entity to expand a hop. Search to jump. A small sample of the graph, running in your browser; what's in it is listed below.
Which Babylon did you mean?
The sample holds several entities called Babylon: the ancient city, a town on Long Island, a 2022 film, an album, a piece of software. Ask a language model about “Babylon” and it picks one and sounds confident about it. Ask a graph with durable identifiers and it hands you all of them and lets you say which you meant.
In the explorer above, open one and the panel offers the others, each labeled by what it sits next to, because that is what tells them apart.
What's in this sample
This page runs on a sample, not the graph. Here is what went in and what didn't.
899 entities, out of 1,063,614
Seeded around Babylon, Mesopotamia and the subjects of the stories, then grown outward.
Mostly from Wikidata, plus a few read by a model and checked by hand
A small number of facts were read out of Wikipedia by a model, and each was checked by hand before it went in; they are drawn in amber. The rest of what the model read stays out.
People are limited to a reviewed set
The graph holds far more people than this. The sample carries only historical figures a human has reviewed, so it under-represents that part of the graph on purpose.
No descriptions, no vectors
Labels and facts only. “Nearest by meaning” works from a small precomputed list. The embeddings themselves stay out.