Multilingual RAG means asking a question in one language and retrieving the answer from documents written in another — without translating the corpus first. In Switzerland this is not a nice-to-have: the same case file routinely contains a German contract, French correspondence, an Italian invoice and an English policy. A retrieval system that only works within one language finds a fraction of the evidence and answers with quiet confidence anyway.
Why translate-then-search fails
The intuitive fix — translate everything into one pivot language and index that — breaks in three ways. Translation is lossy exactly where precision matters, in legal and technical terms. It doubles the corpus and the ingestion cost. And it destroys citability: the passage you show the user is a machine translation, not the document of record, which is useless when someone has to act on the original wording.
The alternative is cross-lingual retrieval: index every document in its original language using multilingual embeddings, which place “Kündigungsfrist”, “délai de résiliation” and “notice period” close together in the same vector space. A German question then retrieves the French clause directly, and the citation points at the real document.
What breaks on Swiss corpora specifically
- Entity names change shape across languages. The same company is “Muster AG”, “Muster SA” and “Muster Ltd”; the same body is “Verwaltungsrat”, “conseil d'administration” and “board of directors”. Without entity resolution across languages, a knowledge graph splits one organisation into three unconnected nodes and cross-document questions fail.
- Keyword search does not cross the language line. Hybrid retrieval usually pairs semantic search with keyword matching; the keyword half contributes nothing cross-lingually. The query has to be expanded into the corpus languages for that half to earn its place.
- Swiss usage differs from the standard variety. Swiss High German avoids the eszett, and Swiss legal vocabulary diverges from German and French national usage — “Betreibung”, “Vernehmlassung”, “fiduciaire”. Models trained mostly on national-standard text handle these unevenly.
- Scanned French and Italian documents. Accented characters are where OCR degrades first, and a mis-read accent silently removes a passage from retrieval. OCR quality is a retrieval parameter, not an ingestion detail.
- Answer language is a separate decision from evidence language. A user asking in French expects a French answer citing the German source — not a French answer that quietly paraphrases without showing the original.
How SovraRAG handles it
SovraRAG indexes documents in their original language with multilingual embeddings covering German, French, Italian and English, expands the query into the corpus languages so keyword matching also crosses the language boundary, and resolves entity mentions across languages so the knowledge graph stays connected. Answers are generated in the language of the question and cite the source passage as written — the reader sees the German clause under a French answer, which is what an auditor or a lawyer needs.
All of it runs inside Switzerland: multilingual retrieval is not a reason to send Swiss corpora to a foreign model, which is the point of a Swiss RAG platform in the first place.
Testing multilingual retrieval before you buy
- Ask a question in German whose answer only exists in a French document. Then reverse it.
- Ask for a comparison across documents in three languages and check whether all three are cited.
- Include a scanned French document with accents and confirm its content is retrievable at all.
- Check whether entity names in different languages resolve to one entity, or fragment the results.
These four tests separate genuine cross-lingual retrieval from an English pipeline with a translation layer bolted on. They belong in the pilot described in enterprise RAG in Switzerland, alongside the sovereignty questions from the evaluation checklist.