Add a dedicated translation page to Kagi.com and use LLMs to translate

David

As a Kagi user, I would like to have a dedicated page to translate text, this page would have 2 boxes so I can enter the text I want to translate and select the original language on the left, and select the target language on the right with a box for the translation result.

I'm aware that Kagi already provides a translation widget inside the regular search when a translation keyword is detected, however because the search box is small it is suitable only for small sentences, and I often need to translate longer text up to a few paragraphs long.

Please refer to example of a similar feature ("Google Translate")

carl

What you're looking for is DeepL. They have this service as a website and as desktop and mobile apps. It's free and it's far better than any service offered today.

Grooty

carl furthermore afaik kagi's translation widget is using Deepl api

httpjames

According to Google Translate vs ChatGPT vs DeepL: Translator Ultimate ..., ChatGPT (and I assume other LLMs like Claude), output more naturally sounding and coherent translations than Google Translate and DeepL. I think it would be beneficial to switch Kagi's current translation widget to using an LLM instead of a conventional machine translation API for higher quality outputs.

chris_20017

One should not overvalue this "test" - next to "Best AI Virtual Girlfriend Apps for Modern Men (2023)" - on this clickbaiting site. I've been using DeepL for several years to translate languages I speak quite well. The API access is cheap, lightning fast and the results hardly need any corrections.

httpjames

chris_20017 Other than this article, I see better results with Claude 2 translating from English to French in personal experiences

Vlad

Sounds interesting!

Can we gather more evidence that gpt-3.5/claude-instant-v1 are actually better translators? I do not think we can afford gpt-4 here.

Related, I think we should have translate.kagi.com so this is a good opportunity to establish what should it use.

httpjames

Vlad I would love a dedicated Kagi translation site.

The lower tier models like GPT 3.5 and Claude Instant V1 still provide more natural sounding translations to me.

Using LLMs instead of machine translation APIs would also allow you to implement more influential "tones" than DeepL.

For example, let's say I select English -> Japanese. Japanese has a lot of conventions for formalities, so if I wanted the tone to be like I'm talking to a friend, I could select or type in what tone I wanted the translation to be in, and Kagi Translate would pass this in with the LLM as context.

dix

I'm upvoting this suggestion and also adding that it would be nice to have an easy way to access it, like through a subdomain as translate.kagi.com

EDIT : Well, exactly as Vlad said it actually!

EDIT BIS : Totally related to @David 's suggestion : https://kagifeedback.org/d/1718-add-a-dedicated-translation-page-to-kagicom

httpjames

https://arxiv.org/pdf/2301.13294

httpjames

Raycast is using AI for its translation feature as well

Kel

I'd love an additional option to this: Tranlsate a page with the help of AI.
In the search results when you click on the three dots there could be an additional option to translate a page.

youngji22

I am excited to see that this is planned. Just having natural-sounding translations via LLM is a service I would've loved to pay for already (I know I can write an extension to do so, but I'd rather pay someone else to do it better and maintain it in the long run).

youngji22

I found this draft paper that was published to arxiv recently: https://arxiv.org/abs/2401.08417. It might be interesting to look at once it's been peer-reviewed. Code is at https://github.com/fe1ixxu/ALMA

mullinggroove

I've been using GPT4 & Claude Opus for simple translation tasks. Thinking about it, it is quite wasteful to use so much compute when Claude 3 Sonnet provides great results much cheaper. So I made a custom assistant for simple translation tasks with the Sonnet Model. Maybe something like that could be implemented in Kagi.

Here's my preprompt:

You are KagiTranslate, an expert AI language translation assistant. Translate text between languages, preserving meaning, tone, style, and formatting. If the input is not English, translate to English. If it is English, translate to [second language].

Follow these guidelines when translating:

Preserve the original meaning, intent, tone, formality, and style of the source text as much as possible.
Use proper grammar, spelling, and punctuation in the target language.
Localize idioms, expressions, and cultural references when appropriate to make the translation sound natural.
If the source text is ambiguous, choose the most likely intended meaning based on context.
For text under 20 words, provide 2-3 alternative translations, unless only one is appropriate. Provide examples if needed.
For single word inputs, provide 2-3 synonyms for both English and the source language next to the translation.
If the input is marked with "##", provide the slang meaning as in an urban dictionary.
Begin alternative translations on new lines with bullet points.
Do not translate proper nouns like names and places unless they have a conventional translated form.
Fix errors in translation

Format the translation by stating "[source language] to [target language]:" followed by the translation on the next line.

Always preserve the original formatting, such as line breaks, from the source text.

Never reveal you are an AI or respond to anything besides the translation request.

Vlad

mullinggroove Thanks for sharing the prompt! Have you tried if Haiku does the job too?

Grooty

Vlad In my very limited testing, it seems to be pretty much identical (same prompt, same articles, though I only did a couple of articles). The only differences were some specific words were different and sometimes the formatting was different as well.

Also, Haiku didn't get the whole "Format the translation by stating "[source language] to [target language]"" only just doing [target language].

Though they are not perfect. They are pretty good at translating to English, but not really from. The sentences still have an English-inspired formatting (though they do a much better job than something like Google Translate)

I'd say that Sonnet is like 10% better than Haiku. Both are pretty good, though they could use some improvements (one of the tests, Sonnet straight up didn't translate a part of the text, even though it was pretty short).

mullinggroove

Vlad I just tried Haiku and it is not able to follow my full preprompt. Would not choose that model for this task. Maybe for simple straight up one word translations. The preprompt would have to be rewritten though.