Background

Actionbase originated at Kakao. For document translation, we plan to use Kanana-2, Kakao's own open-source AI model (kakao/kanana-2).

The goal is to establish a translation workflow that minimizes human review of AI-translated content.

Kanana-2

Kanana-2 is a general-purpose language model developed by Kakao, supporting 6 languages including Korean, English, Japanese, Chinese, Thai, and Vietnamese. While not specifically designed for translation, its 30%+ improvement in Korean tokenization efficiency makes it well-suited for producing natural Korean output. We plan to leverage this model for document translation tasks.

Task

• Draft glossary file (website/i18n/glossary.json)
  • translate: Technical terms and their corresponding translations
  • preserve: Proper nouns and brand names to keep unchanged
• Draft system prompt (website/i18n/translation-prompt.md)
  • Glossary reference rules
  • Style guide (formal tone, polite endings)
  • MDX format preservation rules (code blocks, frontmatter, component tags, relative links)
• Draft style guide (optional)
• Test translation with sample documents and verify quality
• Refine glossary/prompt based on feedback
• Proceed with full document translation

Done When

• Glossary file and system prompt are finalized
• Sample document translation passes quality review with minimal human intervention
• Translation workflow is documented and reproducible

Notes

• Current Korean docs: website/src/content/docs/ko/
• Reference: kakao/kanana-2

If you have any additional action items that would be helpful, please feel free to suggest them.