BGE-M3 Embeddings

bge-m3 is a multilingual embedding model from BAAI. This Replicate deployment returns both dense and sparse vectors from the same input text.

Use it for:

• Semantic search
• Hybrid dense + sparse retrieval
• RAG document indexing
• Multilingual search over English, Arabic, and mixed-language text
• Qdrant named-vector collections with dense and sparse vectors

Inputs

text

A single text string to embed.

texts_json

Optional JSON array of strings for batch embedding. If texts_json is provided, it is used instead of text.

max_length

Maximum number of tokens per text. Default: 8192.

Output

For a single input, the response includes:

• embedding.dense: 1024-dimensional dense vector
• embedding.sparse_indices: sparse token ids
• embedding.sparse_values: sparse token weights
• embedding.dense_dim: dense vector size
• embedding.sparse_token_count: number of sparse lexical weights

For batch input, use the embeddings list.

Example Inputs

Single text:

{
  "text": "The buyer shall pay the purchase price on the agreed date."
}

Batch:

{
  "texts_json": "[\"Murabaha payment clause\", \"Shariah governance standard\"]"
}

With max length:

{
  "text": "The contract must comply with the relevant Shariah standards.",
  "max_length": 4096
}

Example Output

{
  "model": "BAAI/bge-m3",
  "embedding_count": 1,
  "embedding": {
    "text_index": 0,
    "text_length": 60,
    "dense": [0.0123, -0.0456],
    "sparse_indices": [101, 2048],
    "sparse_values": [0.42, 0.31],
    "dense_dim": 1024,
    "sparse_token_count": 2
  },
  "embeddings": [
    {
      "text_index": 0,
      "text_length": 60,
      "dense": [0.0123, -0.0456],
      "sparse_indices": [101, 2048],
      "sparse_values": [0.42, 0.31],
      "dense_dim": 1024,
      "sparse_token_count": 2
    }
  ]
}

The real dense array contains 1024 float values. The sparse arrays vary by input text.

Notes

• Dense vectors are suitable for cosine similarity search.
• Sparse vectors are suitable for lexical or keyword-aware retrieval.
• For Qdrant hybrid search, store the dense vector in a dense named vector and the sparse indices/values in a sparse named vector.
• The model is best used for embedding and retrieval, not text generation.