Building a 12-language AI product search on the edge: what actually broke
I run OneFindMe , an AI product-search front end for AliExpress. You describe what you want in plain language — in any of 12 languages — or upload a photo, and it returns the product, similar items, and cheaper alternatives. It runs entirely on a Cloudflare Worker with an LLM doing the language work. This isn't a launch post. It's the three problems that were genuinely hard, the wrong first…
OneFindMe is an AI product-search front end for AliExpress that accepts queries in 12 different languages. The service runs on Cloudflare Workers and uses the Claude Haiku model for query translation and image identification. Initially, the model attempted to translate shopping queries to English, which resulted in irrelevant search results because sellers typically use short, keyword-rich product names rather than fluent prose.
To address this, the prompt was changed to ask the model to return the short product name an AliExpress seller would use.
Another issue was the model creating non-existent category IDs, which could override real search results when used. To fix this, an allow-list of known-good category IDs was introduced, and the keyword search was kept as the default, with category results being an optional refinement.
The third problem was the search latency, which averaged 6-8 seconds due to uncached searches. To improve perceived latency, the UI displayed bestseller results instantly while the main search ran in the background. Additionally, the first time a translated query was searched, its data was pre-cached for quicker subsequent searches.
The article emphasizes the importance of prompting the model with the target system's language, avoiding relying solely on the model's optional enrichments, and optimizing perceived latency over actual latency in search products.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
