BGE-M3 Embeddings

Embedding model guide for Arabic and multilingual retrieval on DCP.

1. What it is

BGE-M3 (`BAAI/bge-m3`) is a multilingual embedding model from BAAI designed for dense, sparse, and multi-vector retrieval modes.

2. What it does

It turns text into vectors for semantic search and RAG retrieval, with broad language support including Arabic.

3. How it compares

Versus general LLMs: lower cost and latency for retrieval tasks.
Versus rerankers: embeddings are the first-stage retriever; rerankers are usually second-stage quality boosters.

4. Best for on DCP

Arabic document search
Hybrid RAG pipelines
Multilingual retrieval APIs

5. Hardware requirements on DCP

DCP floor: `min_vram_gb: 8` (Tier B)
Recommended providers: 8 GB+ GPUs
Template: `arabic-embeddings`

6. How to run on DCP

Choose template `arabic-embeddings`.
Keep `MODEL_ID=BAAI/bge-m3`.
Tune `BATCH_SIZE` and `MAX_SEQ_LEN` from template env vars for your latency/throughput target.

7. Licensing and commercial-use notes

BGE-M3 is distributed under MIT in the model card. MIT is generally commercial-friendly, but still validate data and output compliance in your product domain.

Sources:

https://huggingface.co/BAAI/bge-m3
/home/node/dc1-platform/docker-templates/arabic-embeddings.json
/home/node/dc1-platform/infra/config/arabic-portfolio.json

Docs