The hard part of 12-language search was not translating it, it was writing like an AliExpress seller

A keyword index has a dialect, and it is not the dialect a language model reaches for by default. Sellers title listings like "Women Elegant Evening Party Dress Sexy Backless" ^[6], so the instruction that worked was not about language at all: return the two-to-three-word noun phrase a seller would put in a title ^[7]. The model's best English was the wrong English, and no amount of quality in the translation would have fixed a query that the index cannot match.

That reframing is also why the model choice lands where it does. If the required output is a short noun phrase and an image label, the task has a low ceiling, and the extra headroom of a bigger model has nowhere to go while still billing on every request. OneFindMe's author reports trying Sonnet and settling on Haiku as the better fit for short product names and image labels, and far cheaper given that every search is an LLM call ^[3]. No per-search price or volume is published in the post, so the cost side of that judgement is his arithmetic, not one a reader can check.

The category-ID failure is the more portable lesson. A hallucinated AliExpress category ID does not raise an error; it returns an empty or garbage result set, which then displaces the keyword-only results the same query would have produced ^[8]. That is a silent regression dressed as a successful response. Two fixes went in, an allow-list of known-good IDs and a rule that keyword search always runs with the category as an optional layer on top ^[9]. The allow-list is hygiene. The layering is the invariant, because it holds even when the allow-list goes stale.

Then the numbers. A cold search does an LLM call, the marketplace round trip, ranking and filtering, and takes 6 to 8 seconds ^[10]; a warm one comes out of Workers KV in about 200 ms, with results cached up to 30 days ^[11]. That is a 30x to 40x spread between the two paths ^[15], which explains why the mitigations are presentational: a bestseller row fills the screen while the real search runs ^[13], and a first-time keyword is saved and pre-cached for whoever searches it next ^[14]. Neither shortens the 6 to 8 seconds. They change who is looking at an empty screen during it.

The claim doing the most work in the whole write-up is that latency on the first search, not relevance, was the biggest driver of bounce ^[12]. It is stated without a bounce figure, a sample size, or a before-and-after, and it comes from the operator's own account of his own product. It is plausible, and it is the sort of thing that gets repeated as a rule once it has been read a few times. Treat it as one shop's experience until someone publishes the curve.