What Is an Embedding?
Turn text into coordinates and you can measure "similar"

① Getting started — a concept we passed three times and never explained

Embeddings showed up in how LLMs work as the step that turns token numbers into "coordinates," in RAG explained as the step that searches document chunks, and in multimodal AI as the step that puts images and audio into the same space. All three times it went by as the single word "coordinates." This piece unpacks that word from scratch. Once you have it, why AI search sometimes fetches an off-target document, and why switching models means reprocessing your data, each take one line to explain.

② Why it's needed — a computer only knows whether letters match

Traditional keyword search checks whether the letters match. So searching "vacation policy" won't find a document titled "annual leave guidelines." A person knows they're about the same thing, but not a single word overlaps.

Converting to token numbers doesn't fix it either. Knowing "apple" is #4823 and "pear" is #91 tells you nothing about both being fruit. A number is just a label. What's needed is a representation where similar meaning gives similar values, and that is an embedding.

③ Understanding it as coordinates

Think of a map. Seoul and Suwon are close, Seoul and Busan are far, because two values (latitude and longitude) are enough to say "close" in numbers. An embedding applies that to text. Each piece of text gets coordinates so that text with similar meaning sits in nearby spots and different text sits far apart.

④ How it gets made — a by-product of training

No one assigns "apple scores 0.9 on the fruit axis." The model reads vast amounts of text and adjusts itself so that expressions that often appear in the same contexts get similar coordinates. An early milestone for this approach was Mikolov and colleagues' 2013 work. When words were turned into vectors, vector("king") − vector("man") + vector("woman") came out closest to vector("queen") — evidence that meaning relationships are encoded as directions between coordinates (Mikolov, Yih and Zweig, 2013).

That example is a symbol of "structure emerges," not a claim that it works for everything. The embeddings used in search today come from far larger, more sophisticated models and are often built per sentence or paragraph rather than per word.

Two kinds of embedding to keep apart

Both use the same idea, "meaning as coordinates," but they are different coordinates from different models. That's also why, in RAG, the embedding model turning documents into coordinates and the LLM writing the answer run as separate steps.

⑤ How "close" is measured — and what it's used for

How close two coordinates are is usually measured by cosine similarity. It looks at how similar the directions the two vectors point, so a value near 1 means the meanings are alike. You don't need to memorize the formula — just remember "the more the direction matches, the more similar."

⑥ Four limits — it only measures "similar"

• ① Similar is not correct. In meaning space, "similar" sentences are sentences on the same topic. Opposite meanings can land close together if the topic is the same. As an illustrative example, "EPS beat expectations this quarter" and "EPS missed expectations this quarter" differ by a single word but can come out very close as embeddings. That's why search can bring back "related-looking" text rather than "the right evidence." It's also why you should open the source and find the sentence yourself.
• ② Every model has its own coordinate system. Coordinates from different embedding models can't be mixed and compared. So if you change embedding models, you must convert all your stored documents again with the new one.
• ③ It can be weak on exact expressions. It's strong at finding text with similar meaning, but for searches that must match the letters exactly — ticker codes, proper nouns, precise figures — keyword search does better. That's why in practice the two are often used together.
• ④ It reflects the tendencies of its training data. Embeddings come from training data too, so bias seeps in (where AI bias comes from).

⑦ Investor angle — coordinates take memory too

An embedding is ultimately a list of numbers, and numbers have to be stored in memory. Here's a simple calculation for a feel. Store 1 million document chunks as 1,536-dimensional vectors, with each number at 32 bits (4 bytes), and you get 1,000,000 × 1,536 × 4 bytes ≈ 6.1GB. Ten times the documents means ten times that (compression techniques can cut it, but the baseline scale is the point).

The more companies attach RAG to their internal documents, the more this kind of storage and search demand grows, and that demand flows through to data-center memory and compute. How that shows up in earnings on the memory-chip side is covered in Micron's FY2026 Q4 results, dissected and the semiconductor sector outlook for H2 2026. That doesn't mean embeddings alone decide memory demand — read it simply as understanding the structure: the more AI services grow, the more numbers there are to store.

⑧ Practical checklist

※ This piece simplifies the concept of embeddings for a general reader and doesn't guarantee the performance of any particular service or model. The coordinate table and the EPS sentences are made-up illustrations, and the memory figure is simple arithmetic that ignores compression. Investing-related content is not investment advice; investment decisions and responsibility rest with the investor.