We build OmniDoc, a document translation and extraction tool. This post is about our own sites, so take the links as what they are.
Our two free tool sites, translate.omnidoc.top and extract.omnidoc.top, were available in 32 languages. Our main site has 93. Today we closed the gap: 61 new languages, 1,210 pages in total, deployed as two Cloudflare Workers. All of it went through machine translation, and quite a lot of it went wrong first. These are the problems we hit and what fixed them.
The setup
Each language needs 540 strings: the page bodies of both sites, a guide page, a comparison page, 8 SEO titles and a few feature cards. Instead of sending 540 separate requests per language, we build one HTML file per language with one <p id="s0001"> per string and translate the whole file. The engine keeps the tags and their ids, so nothing depends on the order of the output, and context from neighbouring strings helps the translation.
61 languages × 540 strings ran as 61 file jobs. The median language came back with 4 of 540 strings missing; the worst (Urdu) lost 15.
Problem 1: "free" means liberty
The first time we did this (for 31 languages), "free PDF translator" came back in 20 languages with the word for free as in freedom: libera, Volný, Wolna, 自由. For a page whose whole point is "no charge", that is the one word you cannot get wrong.
The fix is boring and works: we reword the English only for the engine. "free" becomes "free-of-charge", "for free" becomes "free of charge". The English pages keep their wording. In the new 61 languages we spot-checked 20 titles and every one uses the no-cost word: gratis, ücretsiz, رایگان, 免費, maksuton, bla ħlas.
Problem 2: "free jobs" are job vacancies
"Free jobs wait in a queue" means tasks on the free plan. Several engines translated it as job openings. Same fix: the engine sees "tasks on the free-of-charge plan" and the word "jobs" never reaches it.
Problem 3: product names get transliterated
Left alone, the engine writes OmniDoc in Katakana or Devanagari and turns "Word" into the local word for word. Before translation, every product name, file format and URL becomes a placeholder like {T0}; afterwards they are put back.
The engine sometimes respells the placeholder itself: {Т0} with a Cyrillic T, full-width braces, square brackets. The restore step normalises those first. If a placeholder is still missing, that string is not trusted.
Problem 4: checks that block the build
Every page generator already checks its output: product names and numbers (2000 pages, 501 pages, 2026) must survive, nothing may be empty, no placeholder may be left over. With 61 new languages a handful of strings were bound to fail.
Rather than fixing them by hand, a heal step runs all the checks, and any failing string in a new language falls back to its English source. 70 strings out of about 33,000 needed that. The rest ship translated.
Problem 5: right-to-left and missing screenshots
The old code set dir="rtl" for ar only. The new set includes Arabic dialects (Iraqi, Egyptian, Moroccan and others), Persian, Hebrew, Urdu and Sindhi, so the check became a set.
Each page shows a screenshot of the product in the page's language. Our main site has localized screenshots for its first 32 languages only, so the other 61 would have shown broken images. They now fall back to the English screenshot.
Problem 6: the bundle got big
The sites are static pages bundled into one Worker each. After the change the extraction site is 25 MB uncompressed and 3.2 MB gzipped, which is over the free plan's 3 MB limit but fine on a paid plan. If you are on the free plan, check wrangler deploy --dry-run before you add 61 languages.
After deploy
- IndexNow for the 793 new URLs (Bing, Yandex and the shared endpoint all accepted them).
- Sitemaps resubmitted in Search Console.
- The main site's 61 language homepages now link to the same-language tool pages instead of the English ones, with matching
hreflang.
What we would tell you
- Translate whole files with stable ids rather than single strings.
- Reword ambiguous English for the engine, not for your readers.
- Protect names with placeholders and expect the placeholders to be mangled.
- Let your existing checks decide what ships, and fall back instead of blocking.
If you want to see the result, the free translator is at translate.omnidoc.top (whole PDF, Word or PowerPoint files, layout kept, up to 2,000 pages per PDF, sign-in required) and structured extraction to Markdown and JSON is at extract.omnidoc.top. Both are free; free jobs wait in a queue.
