Plotting the First Point: What an Embedding Actually Is

Opening Scene

Before a county can be mapped, a surveyor has to do something deceptively simple: pick a real point on the ground — a fence post, a rock outcrop, a doorway — and turn it into a number, or rather a set of numbers, a coordinate that says exactly where that point sits relative to every other point on the map. The rock itself doesn’t change. What changes is that it now has an address in a shared coordinate system, one that lets it be compared, measured against, and found again.

An embedding does this exact same thing for meaning.

In Plain English

An embedding is a list of numbers — a vector — that represents a piece of content (a word, a sentence, an image, a product) as a point in a shared coordinate space, generated by a model trained to place similar things near each other. The content itself doesn’t change; what’s created is an address for it, one that lets it be compared, measured against, and found again, the same way the surveyor’s coordinate does for the rock outcrop.

The Old Way

Before embeddings, representing content for search and comparison relied on much cruder addressing schemes:

Keyword-based representations captured a document as a bag of exact words, similar to a survey that recorded only a plot’s street name, with no sense of where it actually sat relative to its neighbors.
Manually assigned categories and tags required a human to decide, upfront, which bucket a piece of content belonged to, a rigid classification scheme that couldn’t capture a plot’s genuine relationship to plots outside its assigned category.
Exact-match lookups treated two nearly identical pieces of content as completely unrelated the moment their surface wording differed, unable to recognize that two differently worded descriptions might genuinely describe the same nearby place.

This reliance on surface-level, rigid representations is precisely what embeddings were built to move past.

What’s Changing (and Why AI Is the Reason)

Modern embedding models, trained on vast amounts of text, images, or other content, can place genuinely similar items near each other in coordinate space even when their surface form differs completely, capturing meaning rather than just matching exact words. This is the foundational capability this entire series builds on, and it’s a direct product of the deep learning advances behind the current wave of AI.
Embeddings can be generated for nearly any kind of content — text, images, audio, even structured records — placing fundamentally different content types into coordinate systems that can, with the right model, be compared meaningfully against each other. This dramatically broadens what “search” and “similarity” can even mean, well beyond what keyword matching ever supported.
Because embeddings are generated automatically by a model rather than manually assigned by a human, they can be produced at a genuine scale — millions or billions of points plotted — that manual categorization schemes could never practically approach. This scale is what makes the rest of this series’ techniques, from nearest-neighbor search to large vector databases, genuinely necessary.

The Metaphor, Fully Extended

Land-Survey Element	Vector Embedding Concept
A real place on the ground, unchanged by being measured	The original piece of content — text, image, or otherwise
A coordinate assigned to that place in a shared reference system	An embedding vector, a list of numbers representing the content
A survey team trained to measure consistently across an entire county	An embedding model, trained to place content consistently in coordinate space
Two plots recorded under different old names but sitting right next to each other	Two pieces of content with different wording but genuinely similar meaning
A modern survey covering millions of points across an entire region	Embedding generation at the scale modern AI systems require

For Beginners: What to Actually Do

Think of an embedding as an address, not a translation — it doesn’t “mean” anything on its own, it only means something relative to other addresses in the same coordinate system.
Get comfortable with the idea that two pieces of content can be “close” in this coordinate space without sharing a single word in common.
Before comparing two embeddings, confirm they were generated by the same model — coordinates from different survey teams don’t line up.
Notice that this article’s lesson underlies everything else in this series: nearest-neighbor search, indexing, and filtering all assume this coordinate space already exists.

For Practitioners and Leaders: The Deeper Layer

Treat embedding model selection as a foundational architectural decision, not an implementation detail, since it determines the entire coordinate system everything downstream depends on.
Recognize that embeddings unlock similarity-based search and comparison across content types that keyword-based systems structurally could not support.
Budget for the genuine computational cost of generating embeddings at scale, since this is the step that makes everything else in this series possible.
Build institutional literacy around what embeddings actually represent, since misunderstanding them as literal translations rather than relative coordinates leads to real downstream mistakes.

Quick Recap

An embedding is a vector — a list of numbers — representing a piece of content as a point in a shared coordinate space, generated by a trained model.
This directly parallels a land surveyor assigning a coordinate to a real point on the ground, without changing the point itself.
Modern embedding models capture genuine meaning rather than surface wording, work across many content types, and operate at a scale manual categorization never could.
Everything else in this series — similarity, search, indexing, filtering — depends on this same underlying coordinate space.

Where This Fits in the Series

This opening article establishes what an embedding actually is, the coordinate this whole series is built around. Article 2 looks at why nearby plots in that coordinate space genuinely share something meaningful.