BGE Reranker v2 M3

1. What it is

BGE Reranker v2 M3 (`BAAI/bge-reranker-v2-m3`) is a multilingual cross-encoder reranker from BAAI.

2. What it does

It reranks candidate documents after retrieval to improve top-k precision in search and RAG.

3. How it compares

• Versus BGE-M3 embeddings: reranker is slower but higher quality for final ranking.
• Versus using only ANN retrieval: reranker usually improves factual relevance on difficult Arabic queries.

4. Best for on DCP

• Second-stage ranking in Arabic RAG
• Enterprise search relevance tuning
• High-quality retrieval for legal/finance KBs

5. Hardware requirements on DCP

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

6. How to run on DCP

1. Start with template `arabic-reranker`.
2. Use default model `BAAI/bge-reranker-v2-m3`.
3. Adjust `TOP_K` and `MAX_SEQ_LEN` in template env vars based on workload.

7. Licensing and commercial-use notes

BGE reranker model card uses MIT. Commercial usage is usually allowed, but verify your compliance obligations for dataset and downstream application policy.

Sources:

• https://huggingface.co/BAAI/bge-reranker-v2-m3
• /home/node/dc1-platform/docker-templates/arabic-reranker.json
• /home/node/dc1-platform/infra/config/arabic-portfolio.json