The question is in German, the decisive instruction in Arabic. A search for the same letters can hardly recognise the matching paragraph. A multilingual search application instead tries to bring question and meaning together across different wordings. That it finds a similar-sounding text, however, is not yet enough.
The official BGE-M3 model card describes multilingual embeddings as well as dense, sparse and multi-vector retrieval. These capabilities concern the selection of sources. A complete translation and a correct answer are further tasks that must be considered separately.
A shared search space is not yet an answer
In dense retrieval, texts are converted into numerical representations and compared using a similarity measure. Sparse methods use weighted features distributed in a different way. Multi-vector methods can compare a text across several representations. The model card describes several retrieval routes; their mere availability does not guarantee a hit for every wording.
For the reader, something simpler matters first: does the retrieved section deal with the same question, including restriction and scope? A hit about archived documents is not a sufficient basis for a question about the current working folder.
Three knowledge cards instead of a large archive
Create a small collection of your own with fixed identifiers. K01 describes an approval rule for sealed packages. K02 allows openly available material to be taken. K03 names a different subject, such as the archive. Keep the facts deliberately short and unambiguous.
For K01, you can work with versions in German, Arabic and Hindi that have been checked by someone with the relevant expertise. The following sentences, which we wrote ourselves, are a teaching example and need a suitable language review before any publication:
K01-de: Versiegelte Pakete dürfen nur nach Freigabe geöffnet werden.
K01-ar: لا يجوز فتح العبوات المختومة إلا بعد الحصول على موافقة.
K01-hi: सीलबंद पैकेट केवल अनुमति मिलने के बाद खोले जा सकते हैं।The identifier links the meaning, not just the file name. Before the search test, define which card each question should hit. For the German question “Wann darf ich ein versiegeltes Paket öffnen?” (“When may I open a sealed package?”), K01 belongs to the expected basis. K02 is a deliberately close counter-example with a different condition.
Judge the hits first, then the wording
Once a suitable search environment is prepared, first ask the questions without any subsequent answer generation. Note the identifiers of the top hits and the respective original texts. This lets you establish whether the right basis is selected at all.
Check the first hit and then a somewhat larger section of the results list. A relevant document in second place shows a different property from a relevant first hit. With a tiny collection, such differences are directly visible without publishing an artificially precise overall rate.
A small evaluation grid is enough: question identifier, expected card, cards found, source language and any difference in meaning observed. Then change only the language of the question. If you rewrite the corpus at the same time, the new round is no longer the same comparison.
Negations are particularly valuable counter-examples
Add a question that targets an exception. Many shared words can make a factually wrong hit attractive. A section with “must not” or “only after approval” has to keep its meaning. A search for similar content should therefore not be confused with a check for an identical statement.
Only after the hits have been selected do you, if needed, have an answer or translation produced. Keep displaying the original passage. This keeps search errors distinguishable from translation errors. A correct source can be rendered badly; a fluent translation can explain an unsuitable source in an understandable but wrong way.
Display is part of the test too
Arabic text needs the right reading direction; technical identifiers stay in their logical order. Hindi is not Indonesian. These differences should already be correct in the reference data before a model test derives an apparent statement about quality from them.
Getting started with multilingual search therefore consists of a few well-understood meanings and good counter-examples. BGE-M3 provides a documented model for this. Whether it makes your collection usefully searchable is only shown by the actual search with the specific questions, languages and original passages.
Sources and status
Sources as of 25 October 2026. BGE-M3 model card.
