LaBSE Tops Armenian Retrieval, OpenAI Wins EN-RU

The blunt takeaway: for low-resource, non-Latin scripts, the “best” embedding model is the one trained for retrieval on multilingual parallel data, not the newest or most expensive API.

• –LaBSE ranked #1 on retrieval with R@1 0.834 and MRR 0.864, ahead of all paid APIs in the benchmark.
• –OpenAI `text-embedding-3-large` did best on EN↔RU but fell to R@1 0.210 on EN↔HY and RU↔HY, showing poor transfer to Armenian.
• –The post’s strongest point is the alignment-vs-retrieval split: some models look good on mean cosine yet fail at actual nearest-neighbor selection.
• –`e5-large` and `e5-large-v2` are presented as “monolingual traps” for this use case, with inflated cosine and weak retrieval.
• –Cohere `embed-v4.0` is reported to regress versus `embed-multilingual-v3.0` on this task, which is a useful warning against blind model upgrades.