Photo: free stock photography (Unsplash licence) — see imprint
What an embedding actually is
A model turns text into a list of numbers positioned in a space where similar meanings sit close together. 'Leaking pipe joint' and 'compression fitting' land near each other even though they share no words. That is the whole magic, and it is genuinely useful.
Why pure vector search fails in B2B
Search for article number 'M8x40-A2' and a semantic model will helpfully return something conceptually similar. Your buyer does not want similar. They want that exact part. Semantics without exactness is worse than keywords.
Hybrid search is the answer
Run keyword search and vector search together and blend the scores, with exact matches on SKU and EAN always winning. You get typo tolerance and meaning without ever losing the part number.
Start with your zero-result log
Before buying a vector database, read what people searched for and found nothing. If those are mostly typos and synonyms, a synonym list fixes it for free. If they are descriptions of problems, that is when vectors earn their cost.
- Vectors match meaning; they ignore exactness.
- Hybrid search: exact SKU always wins.
- Read the zero-result log before buying anything.
Frequently asked questions
Read your zero-result log before you decide, and you may not have to. If people are searching for misspellings and synonyms of things you sell, a synonym list fixes that for free this week. If they are describing a problem in their own words instead of naming your product, that is where vectors earn their keep. The log tells you which shop you have.
Because it is doing what it was built to do. An embedding places text by meaning, and to a meaning-model a part number is just a string that resembles other strings — so it returns something conceptually close. Your buyer typing that number does not want close. They want that part. Semantics without exactness is worse than plain keywords here.
No — run both and blend the scores. Keyword search gives you exactness, vector search gives you meaning and typo tolerance, and hybrid search gives you both with one rule on top: an exact match on SKU or EAN always wins, no matter what the semantic score says. Replacing keywords wholesale is how shops lose their part numbers.
The plumbing is the short part — embedding a catalogue and wiring up a hybrid query is well-trodden work. The long part is deciding what actually gets embedded. If your product data is thin attributes and a manufacturer blurb, the embeddings inherit that thinness and the search stays bad. You are usually buying a data project wearing a search project's name.
We do this for a living — Shopware, Node.js, React, ERP integration and automation for B2B.
Talk to an engineer