When someone says “our search uses embeddings”, they mean this: terms are turned into numbers so that similar meanings sit close together. Sounds abstract? It isn’t. Explained in 5 minutes, without math.
The classic problem
Imagine a customer searches for “V2A” and your search should find products labelled “A2 stainless steel”. Two completely different words. A classical search comparing words has no chance: for it, “V2A” and “A2 stainless steel” are as different as “banana” and “screwdriver”.
Manual synonym lists help, but you would have to maintain every possible variant: “V2A”, “1.4301”, “rustproof”, “non-rusting”, “austenitic”, “inox”. For 280,000 products across 50 material classes, that becomes inhuman.
The idea behind embeddings
Instead of comparing words, an AI converts each word into a coordinate in a high-dimensional space. You can picture this like a map, just with many more dimensions than two.
This map is built so that similar meanings sit close together:
- “V2A”, “A2 stainless steel”, “1.4301”: three points almost on top of each other
- “Cap nut”, “capped nut”: also close together
- “Banana” and “screwdriver”: far apart
When someone searches “V2A”, the search doesn’t look for the word, but for all products whose coordinates lie nearby. That is how it finds “A2 stainless steel”, even though the word “V2A” appears nowhere there.
Where the magic actually happens
The embeddings are not built by hand. An AI learns them from large volumes of text: product descriptions, datasheets, forums, Wikipedia. It sees: “V2A” and “A2 stainless steel” often appear in the same context (“rustproof”, “food-safe”, “screws”). So they get similar coordinates.
At Eywora, embeddings are additionally fine-tuned for your industry. “M12” means something different in a tool shop than in a vehicle parts shop. We train on your assortment so that your meanings come through.
What embeddings can’t do
Three points worth being honest about:
- They are not a magic machine. If you don’t carry a product, semantic search won’t find it either.
- They don’t replace all classical techniques. Typo tolerance, SKU exact-match and filter logic continue to run classically and are complemented by embeddings, not replaced.
- They need data. For very small assortments (<5,000 items), the training effort is higher than the benefit.
How it works in concrete terms
Simplified into three steps:
- Indexing: each of your products (title, description, attributes) is converted once into an embedding vector and stored.
- Search: when someone searches, the input is also turned into a vector, in real time, under 47 ms.
- Similarity: the system looks in the index for products whose vectors are closest to the search query. Those are the top results.
All the magic sits in step 1 (training good embeddings) and step 3 (finding nearest neighbours quickly and correctly). Among the tools for this are vector databases.
Why you should know this
You don’t need to implement embeddings yourself. But if you, as a shop operator, are evaluating search AI, it doesn’t hurt to understand the concept:
- When a vendor says “semantic search”, they almost always mean embeddings.
- When the demo shows “V2A finds A2 stainless steel”: that’s the embedding promise in action.
- When the demo shows “the search learns from clicks”: that’s the combination of embeddings + re-ranking.
More on the concrete impact in the Eywora search module on the Search detail page.