#HardWords: Embedding

Imagine a boardroom. A vendor is presenting you a quote for "an AI solution on your corporate documentation". At some point they say a sentence like: "We'll use latest-generation embeddings to index your documents in a 1,536-dimension vector space."

You nod. None of the ten people in the room asks what it means. The word "embedding" passes, settles into a seventy-thousand-euro budget, and goes into production.

Six months later, the internal company search engine answers "nothing found" to the query "August vacation", and nobody knows why.

Welcome to the second episode of #HardWords, our series that explains in ten minutes the tech terms vendors sell you without ever telling you what they really are. Today, finally: embeddings.

The metaphor that works

An embedding is the translation of a sentence into coordinates.

Imagine a map. Not of streets — of meaning. In this map, every concept, every sentence, every paragraph, every image, has a position. Vacation sits near holiday. Rome sits near Italy. October sits near November. Father sits near fatherhood (but also, surprisingly, near man and son).

This map doesn't have two dimensions like a street map. It has thousands. Literally — modern models use embeddings of 768, 1,536, or even 3,072 dimensions. Don't try to imagine them: no human can. Just know that in that multidimensional map, every sentence is a point, and the closeness between points corresponds to closeness in meaning.

That's all. An embedding is the coordinate vector of a sentence on that map.

Why it exists

It exists because computers have a problem: they don't understand the meaning of words. They can compare letters, they can count. They don't know if "vacation" and "holiday" are synonyms.

For decades, we tried to solve the problem with vocabularies, grammar rules, ontologies. It worked poorly. Then, around 2013, someone did this thing: instead of teaching the computer what words mean, we made it read billions of texts and told it "find a way to represent them as numbers such that words that appear in similar contexts have similar numbers."

Result: the computer built the multidimensional map by itself. It doesn't know what vacation is. It knows that the word "vacation" statistically reminds it of "holiday", "August", "long weekend", and that's enough.

This is the trick. Everything we call "AI that understands language" — chatbots, semantic search engines, summarizers, translators — runs on top of this idea.

What it's for in your company

In practice you need embeddings every time you have to search by meaning instead of by exact words.

Case one: internal search. Your employees search for "travel expense reimbursement" and don't find the document titled "Travel Expense Policy". Embeddings: the two texts are close in the space of meaning. They find the right one even if the words are different.

Case two: conversational assistant on company documents. A customer asks your website chatbot "do you ship to Sicily?". The chatbot doesn't have a specific page on this. But it has a generic logistics policy. Embeddings: the two things are close. The answer arrives.

Case three: recommendations. You have an e-commerce with three thousand products. A customer buys an item. Embeddings: the products closest to the one purchased (in the space of descriptions, images, tags) get proposed.

Case four: clustering / segmentation. You have two thousand support tickets from the last month. You want to understand the 5 most recurring problems. Embeddings: group nearby tickets, and clusters emerge by themselves without having to write rules.

In all these cases, the "old" alternative was: write rules, maintain them, update them every time the vocabulary changes. Embeddings do all this almost automatically. They cost you computing a vector for each piece of content, and saving it in a vector database (yes, that's another hard word; we'll talk about it later).

When you should ask for them

Once you know what they are, the question to ask in the boardroom changes. Don't ask "do you use latest-generation embeddings?" — it's the wrong question, because anyone will answer yes. Ask:

• Which embedding model do you use? (Names you'll hear: text-embedding-3-large from OpenAI, voyage-3 from Voyage, embed-v3 from Cohere, open-source models like bge-m3.)
• How many dimensions? (More dimensions = more nuance, but also more cost. For most business cases, 768 to 1,536 is fine.)
• How do you handle updates? (If you add a new document, it needs to be re-embedded. If the embedding model changes, everything must be re-embedded. Who pays? Who monitors?)
• What language was it trained on? (Many embeddings are excellent on English and mediocre on Italian. If your documentation is in Italian, this is problem number one.)

If the vendor hesitates on one of these questions, they're selling you an idea, not a solution.

When you should be wary

Three red flags, to hang on the wall:

One, someone tells you that "embeddings solve the problem". They don't solve it. They are one layer of the solution. On top there's always something else: how you split the documents (chunking), how you handle context (RAG, which we'll talk about), how you present the result (interface).

Two, they propose to "embed everything and then we'll see". It's the equivalent of "let's dump all our documentation into the database, and the AI will do the rest". Doesn't work. Embeddings without a thoughtful chunking strategy produce random results.

Three, someone tells you that "your embedding will be customized for your company". Almost always false. What's customized is what you embed (your documents), not how (the embedding model is a market standard). Fine-tuning an embedding model is possible but rare, expensive, and makes sense only in very specific cases.

Summary in three lines

An embedding is a translation of a piece of language into a vector of numbers.

Things with similar meaning end up close in this numerical space.

Everything that does "search by meaning" under the hood does this.

Now, in the boardroom, you know what to nod to and what to challenge.

Monday, June 8, we return to the #BraveHistories series, with an Italian case that is an underrated B2B like few: a Marche company of precision measurements that built a silent brand without spending a euro on advertising.

Are they proposing to you "an AI solution on your company documents"? Under artificial intelligence for business we also audit technical quotes, in half a day, free until the end of June. Let's talk.