Multilingual embeddings
Meaning that crosses languages.
A tokenizer splits text; an embedding turns it into a vector so meanings can be compared. This runs Quanfire.ai’s production adapter next to the raw base it was trained over — the same two models we serve — so you can verify the difference yourself, not take our word for it. Type a query and watch it retrieve the same idea across every language.
Search once, match every language
Type a sentence in any language. Both models rank the same fixed corpus — one parallel FLORES-200 slice where every idea exists in all 15 languages. A good multilingual model retrieves by meaning: it should pull the matching idea back in many languages at once. Each colour below is one idea — watch the adapter return a single colour across many language tags, while the raw base clings to surface form.
Embedded live on our GPU · ⌘/Ctrl + Enter to search
Separation, not just similarity
match vs. noise, under each modelA retriever is only useful if it can tell things apart. So we score one anchor sentence twice under each model — against a genuine match, and against an unrelated sentence — and measure the gap between them. The raw base rates almost everything highly similar, so its gap is tiny; the adapter spreads meaning out, so its gap is wide. Bigger gap = sharper retrieval.
…each also scored against a fixed unrelated reference: “The volcano erupted at dawn, throwing ash for miles across the valley.”
The receipt — held-out numbers
recall@1 · higher is betterA small Indic-trained adapter lifts the retrieval it was trained for — and, with no training on them, improves every global language too.
How this stays honest
The base column isn’t a re-implementation — it’s the exact intfloat/multilingual-e5-small checkpoint served through the identical stack as the adapter, with an untrained (zero-effect) adapter attached so the only thing that changes between the two columns is the trained weights. The search corpus is a fixed parallel slice of FLORES-200 devtest — held out from training — and the numbers in the receipt were measured on it. Nothing on this page is cached or hand-picked: every score is computed the moment you press the button.