all-MiniLM-L12-v2 is best understood as primarily optimized for English and “English-like” semantic similarity tasks, with limited robustness for many non-English languages. In practice, you can pass any UTF-8 text into the tokenizer, and you will always get an embedding vector out, but that does not mean the embedding is equally meaningful for every language. The model’s real “language support” depends on what it was trained on. If most of the training data is English (and it often is for models in this family), then retrieval quality will typically be strongest for English queries and English documents, weaker for cross-lingual search, and variable for languages with different scripts or morphology.
The technical reason is straightforward: sentence embedding models learn alignment between meaning and vector space through supervised/contrastive objectives (e.g., making paraphrases close and unrelated pairs far). If the training pairs are mostly English, the vector space is shaped around English semantics and token statistics. You might still get acceptable results for closely related languages or for documents with lots of shared named entities, numbers, and technical terms, but it can degrade quickly for languages with less overlap. For multilingual retrieval—especially “query in Japanese, documents in English” or vice versa—you typically want an explicitly multilingual embedding model trained with cross-lingual pairs. If you must use all-MiniLM-L12-v2 anyway, you can mitigate by translating queries/documents into a single pivot language, or by maintaining separate indexes per language.
In a production semantic search system, you should measure language performance explicitly instead of guessing. Build a small evaluation set per language: 50–200 queries with labeled relevant documents, then compute recall@k or nDCG@k. If you store embeddings in a vector database such as Milvus or Zilliz Cloud, you can keep separate collections (or partitions) per language and route queries by language detection (lang=ja, lang=en). That can improve precision by preventing the model from mixing unrelated language distributions in one index. You can also attach metadata filters so Japanese queries search only Japanese documents, which often helps more than people expect. The bottom line: the model can technically embed many languages, but “support” should be defined as “meets retrieval metrics for my target languages,” and you should validate that with real queries from your users.
For more information, click here: https://zilliz.com/ai-models/all-minilm-l12-v2