What are embeddings? A map with similar things close
Turn a piece of text into a list of numbers that captures its meaning and you have an embedding: two texts about the same thing get similar numbers. So what are embeddings, minus the formulas? Coordinates on a map where similar things sit close together.
What are embeddings? The map of a wine shop
Second Pour, a wine shop in Asheville, North Carolina, sells wine, hosts tastings and ships orders. Picture every phrase on its website pinned to a map: “young red” lands next to “new vintage”; “broken bottle” sits far away, beside “refund.” Nobody sorted them by hand: a model read each phrase and gave it coordinates. Those coordinates are its embedding.
What are embeddings in AI used for? Searching by meaning
For finding something without hitting on the exact word. A customer types “Got anything fruity and cheap?” No listing says “fruity” or “cheap,” but on the map the question lands next to “young red” and “monthly deal.”
What are vector embeddings, and what is a vector database?
“Vector” is the math word for a list of numbers, so “vector embeddings” and “embeddings” are the same thing. A vector database is the warehouse for those coordinates. An ordinary database answers “give me the rows that say X”; a vector database answers “give me the ten texts closest to this one.”
What are embeddings in RAG?
The search step, in most cases. A RAG system looks in your documents before it answers, and embeddings are a common way of finding the passages worth handing to the model. Not the only way: some systems search by words, and many combine both.
What tends to get mixed up
Embeddings and tokens: tokens are the pieces a text is cut into so it can be read; an embedding is the point that sums up what it means. Embeddings and fine-tuning: creating embeddings of your documents doesn’t change the model or teach it anything. It only arranges them so they can be found. And embeddings and the LLM: the model that draws the map usually isn’t the one that writes the answer.
Does my company need this?
You’ll almost never be the one deciding: the tool decides. If it shows up in a proposal, ask what problem it solves. And remember two limits. “Close” isn’t “correct”: the search can bring back something similar and wrong. And with exact references, like a product code, searching by words gets it right more often.
What does that map look like? Tap a phrase
Search for the customer’s question both ways. Then tap any phrase and see which ones sit closest to it.
A customer typesGot anything fruity and cheap?
Twenty phrases from a wine shop’s website, placed by similarity. The ones about the same thing end up together.
Today, Flipa searches by words. So it helps if your documents use the words your team asks with.
This is a drawing to get the idea across, and a simplification: the positions were set by hand, in two dimensions, not computed by a model. A real embedding has hundreds or thousands of numbers, not two.
Search by words or by meaning?
| By words | By meaning | |
|---|---|---|
| What it finds | Texts that contain the words in the question | Texts about the same thing, even in other words |
| When it fails | With synonyms and other ways of saying it | With proper names, codes and exact references |
| What it asks of your documents | That they use the words people ask with | That each topic is explained, not only named |
Many systems combine the two.
How it looks in Flipa
Today, Flipa searches by words, not by meaning. In practice, it finds best what’s written in the words of the question: if your rate sheet says “freight” and the team asks about “shipping costs,” write both. It also doesn’t find words that only appear in a file name or in a web address. Where each answer comes from is explained in how it works. The app’s interface is in Spanish today.
Embeddings and vector databases, no formulas
What are embeddings in AI?
Numerical representations of meaning. A model turns each text into a long list of numbers, and texts with similar meaning end up with similar numbers. That’s how you can measure how alike two sentences are even when they share no words.
What are vector embeddings?
The same thing under its full name. A vector is an ordered list of numbers, and an embedding is a vector that stands for a text, an image or another piece of content. People tend to say “vector embeddings” when the talk turns to storing and searching them.
What is a vector database, and what is it for?
It’s a database built to store embeddings and return, very fast, the ones closest to a given one. It’s for searching by meaning across thousands of texts and for handing an assistant the passages it needs before it answers.
What are embeddings in RAG?
The usual search step. Documents are cut into passages and each passage gets an embedding; when a question comes in, it gets one too, and the system retrieves the nearest passages for the model to read before answering. Some RAG systems search by words instead, or combine both.
Does my company need a vector database?
As a decision of your own, almost never: it’s an internal part chosen by whoever builds the tool. Your job is to check the result. Ask twenty real questions from your team and see whether the right document comes up.
What’s the difference between searching by words and by meaning?
By words, you find the texts that contain the terms in the question: it’s exact, but it doesn’t get synonyms. By meaning, you find texts that are close in sense even if they’re worded differently: it’s more flexible, and less exact with names and codes.
Do your documents use the words your team asks with?
It’s the first thing worth checking, whatever tool you use.
Start free