A new low-rank adapter lets multilingual search tools actually find documents when the query is in a different language than the index.
Researchers built SampoTron, a lightweight LoRA adapter paired with the Nemotron-3-Embed-1B embedding model, that only modifies how queries are read - not how documents are stored. A deterministic router checks what language a query and its target index are in: if they match, the system uses the original encoder; if they do not, it switches to the adapter. Tested across six English, Finnish, and Swedish retrieval directions on a sampled financial benchmark, the adapter lifted average nDCG@10 from 0.241 to 0.291, a 20.7% relative gain. All six cross-language directions improved, and same-language performance, including two full-corpus Finnish evaluations, stayed intact.
The real trick here is what did not change: the document embeddings. Most fixes for cross-language search mean re-embedding an entire corpus in a new model, which is expensive and disruptive for anyone running a live index. This approach instead patches only the query side, making it a plausible retrofit for smaller-language markets like Finnish and Swedish that rarely get dedicated attention from large multilingual models built around English, Chinese, and Spanish.
It is one paper on one financial-domain benchmark, not a shipped product, so treat the 20.7% figure as a promising lab result rather than a guarantee it holds across legal documents, support tickets, or anything outside finance.