Why Zomony Searches in Thai and English at Once

Ask Zomony's AI vision to identify a desk lamp, and it can tell you that in clean English without any trouble -- "desk lamp" is an easy label. The problem is what happens next: Shopee's own product catalog in Thailand is titled mostly in Thai. Searching an English label against a Thai-titled catalog is a mismatch most visual search tools never notice, because most of them stop at "here's what this is" and never get to "here's where to actually buy it."

The gap, found by testing it directly

We tested this with a real product already in Zomony's catalog: a wooden office desk. Searching "Wooden office desk" -- the AI's own English label -- returned zero real matches. Searching "โต๊ะทำงานไม้" -- the same product, in Thai -- matched instantly, at a cosine distance of 0.323 (well inside our similarity threshold). Same product. Same catalog. The only difference was which language asked the question.

That's not a small edge case. It's the default outcome for almost any generic, descriptive search on a mostly-Thai catalog.

Why not just switch to Thai entirely?

We tried that too, and it broke something else: brand names. Shopee's Thai-language titles still carry brand names in Latin script, verbatim -- "Dell UltraSharp" stays "Dell UltraSharp" whether the rest of the title is in Thai or English. Searching a branded product in Thai measured a worse match than searching in English for the exact same product (a Dell monitor scored 0.320 in English against 0.439 in Thai) -- worse being a larger cosine distance, meaning a weaker match. A generic Thai object description, run through a Thai title, sometimes even matched the wrong product entirely, like a search for a well-known mouse landing on a carrying case for it instead of the mouse itself.

So neither language alone is right. English wins on branded products. Thai wins on generic, descriptive ones. The fix isn't picking a side -- it's asking both questions and keeping whichever answer is actually closer.

What Zomony does about it

Zomony's vision model already generates a label in both languages during the same identification pass -- no second AI call, no added cost. When it searches Shopee's catalog for a match, it runs the search in both languages independently, scores each candidate against its own language's embedding (an English label is never compared against a Thai vector, or the similarity score would mean nothing), and keeps whichever match is genuinely closer.

Measured on our own test set: doing this recovered real matches for object labels that returned nothing at all in English alone -- with zero regressions on the branded searches that were already working correctly. Thai does the descriptive-object work; English still wins the branded cases it was already winning. Neither language pays for the other's mistake.

Why this actually matters for shopping in Thailand

If you've ever tried searching a Thai marketplace using an English description and come up empty -- not because the product doesn't exist, but because your words and the listing's words never lined up -- this is exactly that problem. Most product discovery tools are built once, in one language, and treated as "done." A catalog that's genuinely bilingual, the way real shoppers and real sellers actually write, needs a search step that's bilingual too.

Try it yourself

Upload a photo with a mix of branded and generic items on it -- a name-brand keyboard next to an unbranded desk lamp is a good test case -- and see both get matched correctly in the same pass. The same multi-item detection that finds every product in a photo now searches for each one in whichever language actually finds it, and it's free to try.