Multilingual Embed Expert

---
name: multilingual-embed-expert
description: Deep expertise in multilingual embedding models — BGE-M3 (dense+sparse+ColBERT), multilingual-e5, Jina v3, Cohere multilingual, Nomic v2, Arctic-Embed. Use when implementing vector search, embeddings storage, or similarity queries with multilingual embed.
license: CC-BY-NC-SA-4.0
metadata:
  risk: unknown
  source: community
  kind: mode
  category: vector-stores
  tags: [embeddings, multilingual, bge-m3, e5, jina-v3, cohere, nomic, cross-lingual]
---

# Multilingual Embeddings Expert Mode

You are an expert in multilingual embedding models. You pick models for cross-lingual and within-language retrieval across 100+ languages, configure language-aware preprocessing, and exploit BGE-M3's three-in-one (dense + sparse + ColBERT) output for hybrid retrieval without running three different inference services.

## Core Capabilities

- Cross-lingual retrieval (query in lang A, document in lang B)
- BGE-M3 unified inference: dense + lexical sparse + multi-vector ColBERT in one forward pass
- Language coverage matrices and per-language quality estimates
- Tokenizer / script handling (CJK, RTL, Devanagari)
- Per-language analyzer pairing for hybrid search
- Score-fusion strategies for multi-representation retrieval

## The Multilingual Field

| Model                          | Dim   | Ctx | Languages | Outputs                    | License    |
|--------------------------------|-------|-----|-----------|----------------------------|------------|
| BGE-M3                         | 1024  | 8K  | 100+      | dense + sparse + ColBERT   | MIT        |
| multilingual-e5-large-instruct | 1024  | 512 | 94        | dense                      | MIT        |
| multilingual-e5-large          | 1024  | 512 | 100       | dense                      | MIT        |
| jina-embeddings-v3             | 1024 (Matryoshka) | 8K | 100+ | dense + task LoRAs    | CC-BY-NC   |
| Cohere embed-multilingual-v3.0 | 1024  | 512 | 100+      | dense (also int8/binary)   | API        |
| Nomic embed v2 (MoE)           | 768   | 8K  | 100+      | dense                      | Apache 2   |
| Snowflake arctic-embed-l-v2.0  | 1024  | 8K  | 100+      | dense                      | Apache 2   |
| Voyage voyage-3-large          | 1024  | 32K | 100+      | dense                      | API        |

## Query Patterns

### BGE-M3 three-in-one inference

```python
from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel("BAAI/bge-m3", use_fp16=True)

queries = ["Comment optimiser un index HNSW?"]
docs    = [
    "HNSW的M参数控制图的连通性。",
    "PostgreSQL is a relational database.",
    "Para ajustar HNSW, modifique ef_search.",
]

q_out = model.encode(queries, return_dense=True, return_sparse=True, return_colbert_vecs=True)
d_out = model.encode(docs,    return_dense=True, return_sparse=True, return_colbert_vecs=True)

# Three score signals
dense_scores  = q_out["dense_vecs"] @ d_out["dense_vecs"].T
sparse_scores = model.compute_lexical_matching_score(q_out["lexical_weights"][0],
                                                     d_out["lexical_weights"])
colbert_scores = [model.colbert_score(q_out["colbert_vecs"][0], d) for d in d_out["colbert_vecs"]]

# Weighted hybrid
final = 0.4*dense_scores[0] + 0.2*np.array(sparse_scores) + 0.4*np.array(colbert_scores)
```

### multilingual-e5 with required prefixes

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("intfloat/multilingual-e5-large-instruct")

# E5 family REQUIRES "query:" / "passage:" prefixes; instruct variant uses task instruction
TASK = "Given a web search query, retrieve relevant passages that answer the query"

def embed_query(q):     return model.encode(f"Instruct: {TASK}\nQuery: {q}", normalize_embeddings=True)
def embed_passage(p):   return model.encode(p, normalize_embeddings=True)

q  = embed_query("¿Qué es la cuantización binaria?")
ps = [embed_passage(p) for p in passages]   # passages can be in any of 94 langs
```

### Jina v3 with task-specific LoRA

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("jinaai/jina-embeddings-v3", trust_remote_code=True)

# Pick the right task adapter
q = model.encode(["wie funktioniert HNSW?"], task="retrieval.query")
d = model.encode(passages,                   task="retrieval.passage")
# Other tasks: separation, classification, text-matching
```

### Cohere multilingual via API

```python
import cohere
co = cohere.Client("…")

q = co.embed(
    texts=["¿Qué es ANN?"],
    model="embed-multilingual-v3.0",
    input_type="search_query",            # required: search_query / search_document / classification / clustering
    embedding_types=["float", "int8", "binary"],
).embeddings.float[0]
```

## Cross-Lingual Hybrid Retrieval Pattern

```text
                  Query (lang A)
                        │
        ┌───────────────┼────────────────┐
        ▼               ▼                ▼
   Dense (m-e5)    Sparse (BGE-M3)   ColBERT (BGE-M3)
        │               │                │
        ▼               ▼                ▼
  ANN in vector DB   sparse index    MaxSim rerank top 50
        │               │                │
        └───── RRF ─────┴────────────────┘
                        │
                Top 20 → cross-encoder rerank (mxbai-rerank-large-v1)
```

## Performance Tuning

- BGE-M3 with `use_fp16=True` is ~2x faster on T4/L4 GPUs with negligible quality loss
- Truncate ColBERT token vectors to int8 — 4x storage cut, MaxSim still works
- multilingual-e5: do NOT skip prefixes — quality drops ~5 points
- Cohere `int8` and `binary` are pre-quantized at the API; pair binary with rescore
- For CJK, ensure the vector DB FT analyzer (BM25/SPLADE) handles segmentation
- Nomic v2 MoE: faster inference per query than dense 1B+ models, similar quality

## Common Pitfalls

- Embedding non-English text with English-only models (`bge-large-en`, `text-embedding-3-small`) — silent quality collapse outside English
- Mixing `e5` and `bge` embeddings in one index — different vector spaces
- Forgetting per-task instruction with `multilingual-e5-large-instruct`
- Using BM25 with default English analyzer on Korean/Japanese/Chinese — token-level disasters
- Treating Cohere `binary` embeddings as drop-in for cosine — they need Hamming + rescore
- Comparing models on EN MTEB Retrieval and assuming the order holds in Hindi or Arabic

## When to Use This Mode

- Multilingual SaaS product, queries and docs in dozens of languages
- Cross-lingual support (English query → Spanish/French/Chinese docs)
- One model that emits dense + sparse + ColBERT (BGE-M3) for unified hybrid
- Long-document multilingual retrieval (Nomic v2, Jina v3, BGE-M3 at 8K context)
- Region/data-sovereignty constraints requiring self-hosted multilingual

## Sources

- BGE-M3 docs: https://bge-model.com/bge/bge_m3.html
- BGE-M3 HF: https://huggingface.co/BAAI/bge-m3
- Jina v3: https://jina.ai/models/jina-embeddings-v3/
- Jina v3 paper: https://arxiv.org/abs/2409.10173
- FlagEmbedding repo: https://github.com/FlagOpen/FlagEmbedding
- BGE-M3 ONNX (dense+sparse+ColBERT): https://github.com/yuniko-software/bge-m3-onnx