Hi hackers, John and Brendan here from Quetzal (https://getquetzal.com) We’re trying to make it as easy as possible to translate software into other languages using LLMs. Here’s a demo: https://www.youtube.com/watch?v=8hUltBUk9sE.
Internationalizing/localizing is hard. You have to comb through your whole app and look for what needs to be translated and manually update your translations every time new stuff gets added. People sometimes hire their own in-house translators, or they send off their translations to humans to be translated, which takes days, weeks, or worse and costs a ton. Internationalization is, of course, super important for user growth and retention, and often neglected until very late, when competitors will make your product for LATAM, APAC, or Europe.
I worked at Slack for three years on the Slack Connect team, building features like Shared Channels and Shared Workspaces, which had millions of users all over the world. Our translation system was slow and painful, often delaying product launches when things were added last-minute, and resulted in context-less and inconsistent translations. Words like "duplicate" could be either an action or a state, depending on if it's in a button or just displaying a status, and words like "Huddles" were translated ten different ways by translators who just wanted to get through as many translations as possible as quickly as possible, hindering adoption and confusing users.
In a parallel universe, but in the same Downtown Oakland apartment, Brendan worked at a retail startup from the founder of Woot, acquired by Amazon, and they brought someone in to manually translate their product into Spanish, which was expensive and took months. Even worse, the translations had to be manually ported over by engineers every time they updated.
Brendan and I have been coding together since fourth grade, starting by making Cydia tweaks on iOS 4, building apps like one of the first Dogecoin wallets (DogeTicker) in ninth grade, and running small businesses fixing computers and making websites. When we realized we both had miserable translation experiences, we decided to work together to make these approaches a thing of the past!
Our solution to these problems is multiple. Firstly, we utilize a custom Babel plugin to traverse the AST and search for user-facing strings that should be translated, eliminating the need to manually comb through the application and gather strings. During builds, we also utilize the AST to provide context to the LLMs on any new text which is then used to give correct translations for different use cases, (e.g. “Duplicate — is it in a button, where it’s a verb, or is in a regular string, where it’s a noun?)
After scanning for strings and wrapping them in functions to display correct translations to the user based on their browser locale, we gather these strings and build time, and submit new strings to be translated and bring in recently translated strings. During translation, we take both the context and similar strings to guarantee consistency. Translation itself using our LLMs only takes a few seconds, rather than days utilizing humans. Of course, we check them afterwards, but LLMs were themselves built for translation and semantic understanding, so they are incredibly good at delivering context-rich, consistent, and great translations better than humans are.
What's great about this solution is that translations are instant and excellent, and you don't have to spend hundreds of hours finding what needs to be translated and passing them off to an external team, you don't have to wait days for bad quality translations, you don't have to maintain a bot to merge completed translations into the codebase, and you don't have to pass context yourself.
We've got a pilot running on our site right now, so if you've got a Next.js project, feel free to try it out using the script on our homepage: https://getquetzal.com. If you've got a project written in some other framework, like vanilla React or React Native, reach out to us ([email protected]), we've got npm packages that support these too. If you'd like to see support for something else entirely, like Swift, please let us know and we'll move it up on our docket.
Let me know what you all think about how software translation is today. Is it hard? What approaches have worked or not? We also want to explore problems like how people reach new markets and find customers, so that’s something we’re thinking a lot about as well. If you’ve got funny bad translation stories, I’d love to hear those too, the more shocking and hilarious the better.
I think that ideally, every project would be setup from scratch in the beginning to use t() tags or something similar to have translation keys so that setting up translations is as easy as swapping out what t() returns (from e.g. a config file or the like).
Of course, we all know that this is very rarely how projects end up getting setup especially in the early stages, and then it's just massive amounts of work to go back and set it up later.
The thing that's the most intriguing to me about what you're describing is automatically setting up translations in the build step where you auto-detect strings to translate. But looking at the site, most of it seems to be focused around the VSCode extension which will just sort of find and replace strings in the source code with t() tags.
Can you talk more about the translations in the build step? Is there a reason you're not talking more about that on the site? (Is it just newer, not very reliable/good, or...)?
The idea that I could just throw something like this into my project, not have t() tags in my source code but still get translations, sounds like magic and I think it would be really neat.
This is a great idea. I have even hit this pain point when developing a healthcare app for hospitals that was primarily used in the United States. There are certain communities, even just within California, where it is common to have patients who only understand Spanish, Mandarin, or Japanese.
Any plans to extend this to iOS/Android development in the future? I assume it would already be easy to integrate this into React Native.
Also, is there a way for me to provide explicit additional context to the `t` function for the translation? Essentially a string that is appended to the LLM input for translation. For example, in Japanese there is often a significant difference between formal and informal language, and it is common to add post-positional particles such as や, が, and の to make titles and labels sound more natural. I see you have addressed many other special cases around numbers/dates/etc, so certain flags like formal/informal, regional dialect, etc may be valuable future additions.
Overall looks really nice and I look forward to trying this the next time the need arises.
I have been using a very similar approach for i18n of my eternally experimental web game https://luduxia.com/reversi/ which has an entertaining build process including this sort of thing. (I come from a game publishing background so have nightmares of when we all sent Excel sheets around motivating these things).
Does your result live update the strings in place if the device locale is changed?
Do you have any method for getting feedback from UI tests? I don’t now, but that is absolutely a feature I was used to previously. We used to OCR off expected areas to ensure things fit etc.
What measures have you taken to prevent things like this https://github.com/microsoft/WSL/issues/7868
Would be cool to see the Quetzal website internationalized itself and see what it looks like in different languages.
I'm gonna be building a web app in the next month or so that's gonna need a ton of translation, but it's gonna be a SvelteKit app. Do you plan on supporting Svelte anytime soon?
Quetzal as a product is inevitable. It will make it even easier to target international markets from day one with a tiny team. Congrats on the launch! You rock! Greetings from section 4D!
Maybe it doesn't recognize my language settings but it'd be nice if your website would be in my local language to show off that it works.
Funny I have a domain very similarly named been always debating on developing, quetal.com. Otherwise cool concept
I honestly don't get the need for an LLM here, the landing page translations feels unnatural and there are some translations that don't really work in the context - how does this actually perform better than DeepL for instance?
No pricing to be found in the header for an AI product (which you'd expect to be on the pricier side) isn't great either.
Congrats on launching. This looks promising
Just in time. Gonna try it right now. Will share feedback.
Hm. Not convincing. Disclaimer: I am the founder of Opral (https://inlang.com/ && https://lix.opral.com/)
The idea of parsing source code to auto inject translations, especially while leveraging machine translations comes up every 2 months.
It’s not solving the problem.
The problem to be solved is change control. Doing translations is (surprise!) cheap compared to controlling changes. Changes referring to the marketing copy changed, the button label changed, a new screen has been added, etc. It needs one system that can track and control changes across apps, translations, files.
If change control is solved, localization boils down to managing CI/CD pipelines.
Interesting approach! I like the idea of using an ast to find translation strings and to generate source strings by an llm from the context. My experience is the same. Having to deal with 6 languages and a fast changing application is impossible for small teams, so I ended up with asimilar approach where developers only edit the source text, and target translations are automatically kept in sync using Deepl. This solves the whole i18n dilemma of outdated/missing translations, by just focussing on the source texts. It would be great to be able to even generate the source texts from the $t tag. I'm slowly turning this into an open source developer tool: https://garage44.org/posts/expressio/expressio-machine-trans...