Nothing typed. Every text is shown, coloured by its topic. Type above, or click a point to search from its text.
What this page shows
The query becomes one more point, drawn as a ring. Its nearest texts are found in all 1,024 dimensions, never on the screen.
From this search's point of view scales every dimension by the size of the query's own value there. Distances, the result order and the picture are then all taken in that scaled space. It is a linear change, and the status line says when it is on.
Each result says where plain distance had it, so you can see what the point of view changed.
Suggestions under the box list the texts this page can place as you type.
Reranker and Distance difference work as before; clicking a point searches from it.
Decided for the spec
The host supplies the function that embeds a query; the page never holds the gateway address or key.
Search among loaded points works with no embedding, as here.
The reranker is optional and comes from the host.
This view is the fourth one, named Focus; search also works in Tumble, Lattice and Wander.
The point-of-view option changes which results are found, not only how they are drawn.
Where the data comes from
The vectors are real model output, shipped inside the page. Nothing is sent anywhere.
Source
124 words and phrases written for this page, 9 groups
Model
Qwen/Qwen3-Embedding-0.6B, run once on a local gateway; no model runs in the page
Dimension
1,024, stored at 8-bit precision
Distance
cosine, in all 1,024 dimensions
Queries
only texts with a stored vector, or words of them (mean of their vectors)
Reranker
a toy: rewards a word shared with the query
Seed
7
View
linear projection onto 3 directions; a shadow of the space, so length on screen is not distance