A four-state 2-D embedding projection in plain SVG: clusters wrapped in trimmed convex-hull blobs with numbered centroid badges, a nearest-mark hover readout that prints the item's own text, a cross-tab that always adds to 100, and a caveat naming the groups whose distances mean nothing.
Preview in your theme
Loading preview…
This component is part of Pro
The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.
status="ready" · 220 prompts from one day on an LLM gateway, embedded at 3,072 dimensions and projected with UMAP. Five HDBSCAN clusters get an outline and a numbered badge at their centroid; the 12 rows the clusterer left as noise stay muted and get neither, because noise has no centre. Point anywhere in the plot to read the nearest prompt in full, or Tab to the rail and use the arrow keys to isolate one cluster and read how its human labels split.
Latency & TTFT is the largest cluster at 31% of 220 projected items.
Point anywhere in the plot to read the nearest item in full, or Tab to it and step through items with the arrow keys. Pick a cluster below — or Tab to the row and use the arrow keys — to isolate it and see how its labels split.
Read inside a blob, never between them. UMAP is nonlinear: inside an outline, two marks sitting close really were neighbours in the 3,072-dimensional space. The distance between two outlines, the direction from one to another and the size of a blob are artefacts of the layout — Latency & TTFT looking nearer Context window than Refusals & safety here is not evidence about the vectors, and refitting with another seed redraws the map. The axes are left unlabelled for the same reason: they carry no unit. The bar at the bottom left is 2.5 projection units, which is what spread inside a blob should be judged against. Trustworthiness 0.94 means about 94% of each item's 2-D neighbours were also its neighbours upstairs.
UMAP projection of 220 items into two dimensions. 5 clusters outlined, 12 not assigned to any. Largest: Latency & TTFT, 68 items, 31%. Gaps between outlines are an artefact of the layout and carry no meaning. Per-group figures follow in the data table.
Cluster
Items
Share
Spread
Outline encloses
Largest label
Latency & TTFT
68
31%
1.99
63 of 68
performance, 81%
Context window
46
21%
1.79
43 of 46
context, 83%
Pricing & quota
39
18%
2.02
35 of 39
billing, 85%
Eval & leaderboard
33
15%
2.03
30 of 33
evaluation, 82%
Refusals & safety
22
10%
1.58
19 of 22
safety, 86%
Unassigned
12
5%
no centre
not outlined
performance, 25%
The same 220 coordinates with colorBy="label" — grouped by the human intent taxonomy instead of the clusterer's output. The outlines get baggier exactly where the two disagree: "performance" totals 70 items, but 15 of them landed outside the Latency & TTFT cluster. That mismatch is the finding; a tidy picture here would mean the clustering told you nothing new.
performance is the largest label at 32% of 220 projected items.
Point anywhere in the plot to read the nearest item in full, or Tab to it and step through items with the arrow keys. Pick a label below — or Tab to the row and use the arrow keys — to isolate it and see how its clusters split.
Read inside a blob, never between them. UMAP is nonlinear: inside an outline, two marks sitting close really were neighbours in the 3,072-dimensional space. The distance between two outlines, the direction from one to another and the size of a blob are artefacts of the layout — performance looking nearer context than safety here is not evidence about the vectors, and refitting with another seed redraws the map. The axes are left unlabelled for the same reason: they carry no unit. The bar at the bottom left is 2.5 projection units, which is what spread inside a blob should be judged against. Trustworthiness 0.94 means about 94% of each item's 2-D neighbours were also its neighbours upstairs.
UMAP projection of 220 items into two dimensions. 5 labels outlined. Largest: performance, 70 items, 32%. Gaps between outlines are an artefact of the layout and carry no meaning. Per-group figures follow in the data table.
Label
Items
Share
Spread
Outline encloses
Largest cluster
evaluation
39
18%
4.42
32 of 39
Eval & leaderboard, 69%
context
48
22%
2.74
43 of 48
Context window, 79%
safety
24
11%
3.29
21 of 24
Refusals & safety, 79%
performance
70
32%
3.62
61 of 70
Latency & TTFT, 79%
billing
39
17%
3.37
33 of 39
Pricing & quota, 85%
status="empty" · a valid contract with no coordinates at all, worded so it cannot be mistaken for a failed fetch — the card draws a projection, it never computes one, so the fix is upstream.
This card draws coordinates, it does not compute them: run UMAP over your vectors upstream and pass the two columns it returns, one row per item.
status="loading" · a skeleton holding the square plot box, three blobs where the clusters will land and the group rail below it, aria-hidden, honouring prefers-reduced-motion, with no timer involved.
status="error" · the reducer job or the request for its output failed. Try again only appears when onRetry is passed, and clicking it flips the card to ready.
The reducer job or the request for its output didn't come back. Nothing is drawn rather than a partial cloud — half a projection has a shape, and that shape would be wrong.
Zyeon UI