How does a material match score?

Pick what a designer typed on a construction row, then change how the search works. The materials are a fixed sample. Every score comes from the real search models, computed ahead of time.

Watch the explainer · 1 min 29 s · how a material match is scored, from the text the designer types to the top-five list.

Designer typed

How the search works

Nothing here is saved to the product. Every change re-runs the real search on the 18 sample fabrics straight away.

Look for the same words Look for the same meaning

Two separate searches run on every query. Switch one off to see what the other finds on its own — the surprises usually come from the meaning one.

A typical word hit is worth

How much one matching word counts on its own. At 50%, one word found on a fabric’s card is worth half a match. Raise it when exact wording matters; lower it when a near-synonym should do.

Words and meaning together

Only show matches above

The lowest score a fabric needs to be offered. Below this line it is dropped — and if nothing clears it, the designer is offered “create a new material” instead.

Show

How many results the designer is shown at once. The API lists the top 5 by default and never more than 25 — anything that clears the cut-off but sits below that line stays hidden.

Best matches

How material best match works, in plain words

The designer types a material name on a construction row, such as “cotton poplin”. That text is the whole search. Brand, tags and images are not used.
An ID wins outright. If the text is exactly a Material ID or Supplier ID in the library (capitals don’t matter), that material is shown at 100% and nothing else.
Otherwise it looks two ways. Same words: a library material is a hit only if it contains every word typed, after trimming endings (“linings” finds “lining”). A typo or one missing word means no hit. Same meaning: an AI model turns the typed text and each material’s card (name, ID, content, construction, weight, supplier) into a fingerprint of its meaning, so “sweatshirt fleece” lands near French Terry even with no shared word.
The two scores combine so a word hit only ever raises a material: 1 − (1 − words) × (1 − meaning). Words at 50% and meaning at 60% give 80%.
Anything under 35% is dropped. If nothing is left, the designer is offered to create a new material instead of a weak match.
Accept or reject is recorded with what was shown, so the cut-off can later be tuned from real choices. Accepting copies the library fields onto the row. Later library edits do not change the row.

What the sample shows

Meaning scores sit low: the best real match for “Cotton Poplin” is 74%, unrelated fabrics 30–45%. That is normal for this model, so 35% lets loosely related fabrics through (“silk chiffon” shows Viscose Satin and Polyester Lining).
Words need every word: “stretch denim” hits only Stretch Denim, “wool coat fabric” hits nothing, because no card says “coat” or “fabric”.
Supplier names are on the card, so “Toray” finds both Toray fabrics.