I'm building a multilingual search feature and need help choosing an embedding strategy

A good multilingual embedding strategy usually depends on one question: do you want one shared semantic space for all languages, or language-specific spaces with translation/query routing?

Recommended default

For most search products, I’d start with one multilingual embedding model for both documents and queries.

Why this is usually best

• Cross-lingual retrieval: a Spanish query can find an English document if the model is truly multilingual.
• Operational simplicity: one index, one vector space, one pipeline.
• Better UX for mixed-language corpora: users can search across languages without needing language detection to reroute.

Good model characteristics

Look for models that:

• support many languages well,
• are trained for retrieval, not just similarity,
• handle short queries and long documents,
• have strong performance on your target languages.

Examples of model families to evaluate:

• bge-m3
• multilingual-e5
• LaBSE
• some commercial embedding APIs with multilingual support

When to use a different strategy

1) Language-specific embeddings per language

Use this if:

• your languages are very different in quality/coverage,
• you have heavy language-specific ranking needs,
• cross-lingual search is not important.

Pros:

• sometimes better monolingual relevance per language,
• can tune each language separately.

Cons:

• more complexity,
• harder cross-lingual search,
• more operational overhead.

2) Translation-based retrieval

Translate query into document language or all docs into a pivot language.

Use this if:

• you need very high relevance,
• your chosen embedding model is weak for some languages,
• you already have translation infrastructure.

Pros:

• can improve recall for low-resource languages.

Cons:

• latency, cost, translation errors,
• can distort named entities and domain terms.

3) Hybrid search: BM25 + embeddings

For production search, this is often the best overall approach.

Use:

• BM25 / keyword search for exact match, names, rare terms
• multilingual embeddings for semantic matching
• optional reranker for final ranking

This is especially strong for multilingual search because:

• embeddings help across languages,
• keyword search preserves exact-term precision,
• reranking improves relevance.

Practical recommendation

If I were designing this, I’d do:

1. Use a multilingual retrieval embedding model
2. Index all documents in one vector index
3. Keep language metadata
4. Combine vector search with keyword search
5. Optionally rerank top results with a multilingual cross-encoder/reranker
6. Evaluate separately for each language and cross-lingual pairs

How to choose the embedding model

Benchmark candidates on:

• same-language search
• cross-language search
• named entities
• domain-specific queries
• short vs long text

Measure:

• Recall@k
• nDCG@k
• MRR
• latency and cost

A simple decision rule

• Need cross-lingual search? → use a multilingual embedding model in a shared space.
• Need best precision and can afford complexity? → add hybrid retrieval + reranking.
• Need strong support for low-resource languages? → consider translation or language-specific tuning.

If you want, I can help you pick a strategy based on:

1. your languages,
2. corpus size,
3. whether queries and docs are in the same language,
4. latency/cost constraints,
5. whether you want open-source or API-based embeddings.