GuidesUnderstanding AI

Multilingual embeddings: search in one language, find in another

By the Banzai Labs teamUpdated on 2 min read

chienFRdogENperroESSame meaning, same areaA question in Spanish finds a document in French
Chien, dog and perro land in the same place in vector space.

The principle

  1. 1One modelFor every language.
  2. 2One shared spaceSame meaning, same area.
  3. 3One searchThe question finds the meaning, not the words.

Use cases

Case Gain
Support in several languages A single knowledge base
Translated store Search understands every language
Mixed internal documentation Employees search in their own language
International monitoring Articles from several countries grouped by topic

Testing properly

  • Per languageA set of questions in each language.
  • Across languagesQuestion in one, document in the other.
  • Domain termsTechnical words are often the most fragile.

The basics: embeddings explained. A case: a multilingual RAG chatbot. Search: hybrid search.

Frequently asked questions

Do documents need translating before indexing?
Not with a multilingual model: each document is indexed in its own language. Translation is still useful for showing the answer in the user’s language.
Are all embedding models multilingual?
No. Some are trained mostly on English. Check the list of supported languages and test on your languages, especially the less common ones.
What about keyword search?
It does not cross languages: “chien” does not find “dog”. In hybrid search, the keyword part mainly helps in the language of the question.

Read next

All guides