DigUp Guide Unofficial
EN

DigUp Cross-Language Search

DigUp enables cross-language search across more than 100 languages by using Google DeepMind's EmbeddingGemma 2 model to match meaning across text, images, and audio.

Current versions: 0.6.1 for Mac, 0.7.0-beta.1 (beta) for Windows and Linux; checked against DigUp's GitHub releases on October 15, 2026.

How Multilingual Search Works

A search query in English can find a note written in Bengali, Arabic, or any of the 100+ languages understood by the model. The reverse also applies: a search written in Arabic or Bengali can retrieve English files.

Google notes that EmbeddingGemma 2 is not equally strong in every language. DigUp’s ranking evaluations cover English, Bengali, and Arabic because those are the languages included in the project’s test files.

Limits on Exact Words

DigUp combines vector similarity with exact-word matching. While cross-language conceptual retrieval works across the model’s supported languages, two limits apply specifically to exact words:

  • Languages without spaces: Languages written without spaces between words, including Chinese, Japanese, and Thai, are not segmented into individual words for exact string boosting.
  • Screenshot text: Text inside screenshots is read only on macOS through Apple’s text recognition, which supports about 25 languages. On Windows and Linux, screenshot text is not read at all.

Separately, version 0.6.1 on macOS resolved an issue where composing characters using an Input Method Editor (IME) for Korean, Japanese, or Chinese caused loose letters to appear in the search bar. That search input fix is distinct from how the model indexes content.

Questions

Can I search in English to find files written in another language?

Yes. DigUp uses EmbeddingGemma 2, which understands more than 100 languages. An English search query can find documents or notes written in Bengali, Arabic, or other languages, and non-English queries can locate English files.

Is DigUp equally strong in every language?

No. Google notes that EmbeddingGemma 2 is not equally strong across all languages. DigUp's own evaluation tests only cover English, Bengali, and Arabic because those were the languages present in the test files.

How does exact matching work for Japanese, Chinese, or Thai?

Exact word matching does not split languages written without spaces, such as Chinese, Japanese, and Thai, into individual words. However, semantic search using the model's vectors still functions for those languages.

Can DigUp recognize text inside screenshots in all languages?

No. Screenshot text recognition is available only on macOS, where Apple's Vision framework supports about 25 languages. DigUp on Windows and Linux does not read text inside screenshots.