Add ChatGPT translator.

[BrightXiaoHan]

Is it possible to adapt openai api to this project?

[vitonsky]

As i know, ChatGPT it's a proprietary language model learned to be chat partner. So ChatGPT is not have API to translate texts.

But if you can adopt their API to translate texts, you can add translator to a Linguist.

Actually, you can add any translator implementation, see custom translators docs.

All you need - implement JS class with few methods that will handle requests to translate

[vitonsky]

I investigated your idea, it looks that it possible to implement.

Also, i found a browser extension https://github.com/yetone/openai-translator that implement your idea.
This browser extension cannot to translate web pages, but can translate single text.

I will implement ChatGPT translator in future release. Thank you for idea

[lrq3000]

I would suggest to support OpenAI API, with a custom URL, so that you would get support for both OpenAI's ChatGPT AND local LLMs with a single function (because most local LLMs implement an "OpenAI-like API" to try to be compatible with apps that already support ChatGPT). Eg, ollama, koboldcpp and text-generation-webui all offer OpenAI-like APIs.

[a0s]

Whats intersting, why this project has chatgpt api integration already, but its not available as sort of example of custom module?

[a0s]

The easies chatgpt api custom translator. It would be nice to have this in linguist package, so to just add the openai key.

class ChatGPTTranslator {
    constructor(model = 'gpt-4o') {
        this.apiKey = 'YOUR-OPENAI-API-TOKEN';
        this.apiUrl = 'https://api.openai.com/v1/chat/completions';
        this.model = model;
    }

    async translate(text, from, to) {
        const prompt = `Translate the following text from ${from} to ${to}:\n\n"${text}"`;
        const response = await fetch(this.apiUrl, {
            method: 'POST',
            headers: {
                'Content-Type': 'application/json',
                'Authorization': `Bearer ${this.apiKey}`
            },
            body: JSON.stringify({
                model: this.model,
                messages: [{ role: 'user', content: prompt }],
                temperature: 0.3
            })
        });
        const data = await response.json();
        return data.choices?.[0]?.message?.content || '';
    }

    async translateBatch(texts, from, to) {
        return Promise.all(texts.map(text => this.translate(text, from, to)));
    }

    getLengthLimit() {
        return 4000;
    }

    getRequestsTimeout() {
        return 1000;
    }

    checkLimitExceeding(text) {
        if (Array.isArray(text)) {
            return text.some(t => t.length > this.getLengthLimit()) ? 1 : 0;
        }
        return text.length > this.getLengthLimit() ? 1 : 0;
    }

    static isSupportedAutoFrom() {
        return true;
    }

    static getSupportedLanguages() {
        return ['en', 'pl', 'ru'];
    }
}

ChatGPTTranslator;

[vitonsky]

@a0s if you have tested this code and it works fine - you can add your code to official repository with custom translators https://github.com/translate-tools/linguist-translators

Just create pull request and i will review your code and merge if it works fine.

Keep in mind that any AI models may respond you irrelevant text, so we have to handle such cases.

The trick i know is we can tell to ChatGPT to send us translations in JSON object with some structure like {"translation": "text"} and then parse response JSON and if fails - repeat our request with limit for attempts.

This trick helps to avoid irrelevant nonsense in translations like a translated text with AI comments/annotations.

[a0s]

@vitonsky Thank you for this hint. In addition, I noticed that when gpt does not know the context (for example, if we translate one word), it starts giving several translation options directly in the response, which is not the expected behavior.

When "Auto detect" is selected as a source in the interface, what do we get in ${from} ?

I see the behavior that when we do the translation "Auto detect" -> "Some language" - the translation occurs but shows an empty string. If I then switch to the source language - the translation appears (apparently from the cache)