Semantic search · Retail & industry
Semantic search for industrial catalogs: 200,000+ SKUs found by meaning
An industrial distributor with more than 200,000 SKUs had the problem the whole sector knows: the search box only found what was typed exactly as it appeared in the product master. We built a search engine that understands what the user means — and shipped it to production: today the right product appears as the first result in 65% of searches, and within the top three in 90%, even when the customer uses another term, another language, or simply describes what the product is for.
Problem
Catalogs with 200,000+ SKUs where exact-word search fails to find what the user actually needs.
Result
The right product is the first result in 65% of searches and in the top three in 90% — even when the user writes another term or another language.
The context
In an industrial catalog, product descriptions are not written by marketing: they are written by whoever registers the product. Abbreviations, manufacturer codes, sizes glued to the text, trade jargon. The result is that a search for "gloves for chemicals" does not find "NITRILE GLOVE GREEN S-9" — even though it is exactly what the customer needs.
The cost of that is silent but enormous: searches ending in "no results," customers depending on the veteran salesperson who knows the catalog by heart, and sales lost because the product did exist — but nobody found it. With over 200,000 SKUs, no human being can be the index.
How it works
The solution was to stop searching for words and start searching for meaning. Every product in the catalog becomes a vector — an embedding — that captures what it is about: material, use, category, technical synonyms. The user query becomes another vector, and the system finds the closest products by meaning, even when they share not a single word. And when the exact product does not exist in the catalog, that same meaning space proposes the closest substitutes — the other half of a sales counter’s real work.
But the embedding is not computed on the raw description. It is cleaned and enriched first: abbreviations normalized, jargon expanded, brand, category and attributes structured. That text curation proved as decisive as the embedding model itself — half the project was data engineering, not AI.
Engineering decisions
Decision 01
Hybrid search, not vectors alone
Vectors find by meaning, but an exact part number is found better by traditional search. We combined both: keywords for the literal (codes, brands, measurements) and semantics for intent. The final ranking blends the two signals.
Decision 02
Curate the text before embedding
We normalized abbreviations ("S-9" → "size 9"), expanded trade jargon and structured attributes. An embedding computed on dirty text returns dirty results: search quality is decided before you ever touch the model.
Decision 03
The catalog changes every day
Additions, removals and price changes sync automatically, and the vector index updates without maintenance windows. A search engine that shows discontinued products destroys in a week the trust it earned in months.
What changes for the business
The deep benefit is not "a better search box": it is that the expert salesperson’s knowledge stops being the only path into the catalog. Anyone — a new customer, a junior rep, a procurement team — finds what they need by describing it in their own words.
- 65% of searches return the exact product as the first result, and 90% within the top three — measured on real searches.
- Fewer "no results" searches: when the exact product is missing, the system proposes the closest substitutes.
- The entire catalog becomes sellable, not just the part the team knows by heart.
- The same semantic foundation is ready for quoting engines and AI agents over the catalog.
Where else it applies
Industrial distributors
Technical catalogs with tens of thousands of SKUs and cryptic descriptions.
B2B e-commerce
Search that converts: the customer describes, the system finds.
Spare parts & automotive
Find the part by function and equipment, not just by code.
Pharma & laboratory
Supplies with technical nomenclature and multiple synonyms.
FAQ
Frequently asked questions about this case
Do I need to replace my ERP or online store to get semantic search?
No. The search engine integrates as an API: your platform sends the query and receives ranked results. The product master keeps living where it lives today; the system syncs with it.
Does it work with Spanish descriptions, abbreviations and jargon?
Yes — that was exactly this case: cryptic descriptions, sizes glued to the text and trade jargon. Part of the work is building the normalization dictionary specific to your catalog.
How long does it take to put semantic search in production?
It depends on the state of your data, but the pattern is proven: a first useful version ships to production in weeks and is then tuned with your users’ real searches.
What happens with new or discontinued products?
Synchronization is automatic: additions, removals and price changes flow into the index every day without manual intervention.
Have a similar challenge?
Tell us your problem. We'll tell you how to solve it — and how long it takes.