Local document search across languages

Different words, the same subject

An employee searches in Polish for „przekazanie obsługi po zakończeniu projektu”, the handover of service after a project ends. The report uses the phrase „service handover”. An ordinary search finds nothing, because the document does not contain a single word from the question. So the employee writes to a colleague in the international team and waits half a day for an answer.

A search that compares meaning points to the right chapter despite the different language. Next to it, though, appears a report about handing over goods, which concerns another matter. The user opens both passages and picks the right one. Terms that sound alike need not mean the same activity, and a company glossary helps fix that distinction for later questions.

Translating the whole archive would be a separate, expensive project. Multilingual search avoids that cost, because the employee reads only the passage found and, if needed, translates just that paragraph.

What multilingual document search costs and when it pays back

The model is a small item here. The result depends on the time that searching and asking around take today, and on how much of it remains for checking the answer. In the option where a model on your own GPU also writes the answer, a GEX45 server at €214 a month comes on top, which extends the payback to 1.9 months.

Before we quote the build, we run a test on 20 to 100 questions the team has already asked. We compare the app's answer with what an employee found by hand and count the time from question to source. The test takes 5 working days and costs €0.

How to test the model before you buy

The BGE-M3 model card describes retrieval in more than 100 languages and texts of up to 8,192 tokens, and the authors recommend combining search by meaning with keyword search and reranking. That claim does not replace a trial on your own language pairs and terminology. The test also covers questions the collection has no answer to: the user should see no results instead of a source forced to fit. A reviewer who knows both languages judges whether the passage meets the need and whether the original opens at the right place.

Where to start

If the material must stay in your own environment, preparing the text for search and any translation of the question also run locally. With IT we agree the server, backups and access to documents, so that a user never sees a source they may not read. We build the search as part of private LLMs and AI apps.

For the first call, one question that keeps coming back to colleagues and a note of where the team looks for answers are enough. You do not need to send confidential documents, and you get a written plan within 2 working days.