Every dot on that map is one meditation from the DailyMedToday archive. Its position comes from the pipeline this series built: each text embedded into 1,024 dimensions, the whole space flattened to a plane with UMAP.
The colors are incidental. The positions are everything.
Look at the structure. An island sitting alone on the left. A tight knot at top center. A cluster in the lower right. A grouping in the middle. These aren’t artifacts of plotting — they’re neighborhoods in meaning-space, groups of meditations that the embedding model placed near each other because of what they say.
Reading the map
So I read the clusters. Hovered over every dot, identified what topics lived where. Here’s what the geometry found:
- The island on the far left? Every meditation in it is about resurrection.
- The small grouping toward the middle? All about heart.
- The red in the lower right? Love
- The topmost cluster? The Unseen.
I didn’t manually tag or label these meditations. There is no “resurrection” category in the database. No editor sorted the archive into themes, and no rule told the system which doctrines exist.
The meditations were written one per day over months, each on its own — and the structure emerged from the texts themselves, recovered by a model that has never been to seminary, from nothing but the way language carries meaning.
This visualization should drive home the full series – it’s the thesis we’ve been focused on since September. Part 1 claimed that embeddings encode meaning as position and that closeness is the signal. This map is evidence of claim: the geometry independently rediscovered the thematic organization a thoughtful human editor would have otherwise imposed.
But the geometry was never directly being asked to.
Reading it honestly
The within-cluster groupings are the trustworthy signal — local neighborhood structure is what UMAP preserves, and “these meditations grouped themselves by theme” is a claim the math supports. The space between clusters is not: the resurrection island’s distance from the repentance cluster is a projection artifact.
It’s tempting to read a measurement of doctrinal distance into the clustering. Such a measurement doesn’t exist.
The clusters are real; the gaps are an artifact of the projection. Trust the neighborhoods, not the ocean.
Why this matters beyond a meditation app
Swap the corpus and the same lesson works with any large body of content.
Point this pipeline at support tickets and the clusters are your real product-problem taxonomy — the one customers experience, not the one in your carefully-curated dropdown. Point it instead at a documentation site and the lonely outliers are orphaned pages while overlapping knots are redundant ones.
Years of meeting transcripts (part 10 made audio searchable) and recurring themes surface that no one was tracking. In every case the embeddings reveal the structure a content collection actually has, as opposed to the structure someone once guessed it should have.
Taxonomy is a hypothesis; the map is evidence.
The validation runs deeper than visualization. The clusters are the proof that the representation underneath is sound — the same geometry that groups resurrection meditations into an island is what makes “overwhelmed by life” find the meditation about rest (part 2), what makes RAG retrieve the right context (part 9), what makes a spoken sentence findable years later.
One representation, audited by your own eyes, powering everything.
Where you go from here
This series this started with a definition: embeddings turn meaning into coordinates. Along the way it became a working stack — ext-infer, ext-turbovec, ext-whisper, ai-contracts, ai-toolkit — all open source, all MIT, all running on hardware you control.
Remember, your data never leaves the machine.
If you build something with it, I want to hear about it. And if your organization is sitting on an archive — text, audio, tickets, transcripts — wondering what structure is hiding in it: Let’s talk.
Discussion