# README
Translate json files with GPT-4/OpenAI
go install https://github.com/kevincolemaninc/i18n-translate-go@latest
or just pull and run the project
$ export OPENAI_API_KEY=...
$ go run main.go -file "~/projects/rn-app/i18n/en.json" -lang vi
output directory: output-{language}.json
flags
| flag | description |
|---|---|
| file | path of the source english file |
| lang | language you want to translate to, (this is sent to gpt-4 and doesn't and shouldn't be a language abbreciation (vietnamese is better than vi) |
| output | destination of the json file. default is output-{lang}.json |
| model | name of the completion model. default is gpt-4-turbo |
| chunksize | To ensure accurate translation and prevent skipping phrases, limit the number of letters translated per request. For common languages, 2000 letters are suitable, while for less common languages like Lao, opt for 500 letters. The default limit is 2000 letters. |
features / roadmap
- concurrency (5 workers)
- support multiple gpt models
- cache results (only update missing keys)
- automatically check for blank or missing translations
- retry blank or missing translations
Example output
with 5 workers, chunksize of 1000, and 26,000 letters, this takes about 2 minutes
$ i18n-translate-go -file "./src/utils/i18n/en.json" -lang "korean" -output ko.json -model gpt-4-turbo -chunksize 1000
File Path: ./src/utils/i18n/en.json
Language: korean
This can take a few minutes b/c GPT4 is slow
Progress: 5/49
Saved result in: ko.json
How does it work?
- Flattens the json into a nested key structure: { "user": { "name": .. } } -> "user.name".
- Chunks the key/values by length of the keys in characters [0].
- Sends the chunks to completions API using function calling.
- Unflattens the resulting json and saves it to disk.
The prompt is here.
[0] - to minimize complexity and dependencies, I count characters instead of tokens
Known errors
Translation error - you should restart this b/c the translations will not be complete.
If the chunkSize is too big and/or the language is in uncommon language (e.g. Lao), chatGPT doesn't translate all the strings you ask it.