BGE Reranker
Basic Information
- Organization: BAAI (Beijing Academy of Artificial Intelligence)
- Country/Region: China (Beijing)
- GitHub: https://github.com/FlagOpen/FlagEmbedding
- Hugging Face: https://huggingface.co/BAAI
- Type: Open-source Reranking Model
- License: MIT License
Product Description
BGE Reranker is a series of open-source reranking models launched by BAAI, as part of the FlagEmbedding project. Unlike BGE embedding models, the reranker takes both the query and document as input and directly outputs a similarity score (rather than embedding vectors). BGE Reranker employs a cross-encoder architecture, performing full attention computation on input pairs, making it more accurate but also more time-consuming than embedding models. Its multilingual version, bge-reranker-v2-m3, is the most recommended choice among open-source reranking models.
Model Matrix
Multilingual Models
| Model | Base Model | Features |
|---|---|---|
| bge-reranker-v2-m3 | - | Preferred for multilingual, excellent in Chinese and English |
| bge-reranker-v2-gemma | gemma-2b | English + multilingual, strong performance |
| bge-reranker-v2.5-gemma2-lightweight | gemma-2-9b | Lightweight, supports token compression and inter-layer lightweight operations |
Chinese-English Models
| Model | Features |
|---|---|
| bge-reranker-v2-m3 | Multilingual general-purpose |
| bge-reranker-v2-minicpm-layerwise | Layer-wise lightweight |
English Models
| Model | Features |
|---|---|
| bge-reranker-large | Large-scale English reranking |
| bge-reranker-base | Basic English reranking |
Core Features
- Cross-Encoder: Full attention computation on query-document pairs, higher accuracy than embedding models
- Multilingual Support: v2-m3 supports reranking in multiple languages
- Layer-wise Lightweight: v2.5-gemma2-lightweight supports token compression and adaptive layer operations
- RetroMAE Pretraining: Pretrained using the RetroMAE method
- Large-scale Contrastive Learning: Trained with contrastive learning on large-scale paired data
- Multi-benchmark Validation: Significant improvements in BEIR, C-MTEB, MIRACL, LlamaIndex evaluations
Business Model
- Completely Free and Open-source: MIT License
- Hugging Face Distribution: All model weights are publicly available on Hugging Face Hub
- Azure AI Integration: Available on the Azure AI platform
- NVIDIA NIM: Deployable via NVIDIA NIM
Target Users
- RAG system developers (especially for Chinese scenarios)
- Teams requiring local deployment of reranking
- Multilingual retrieval system developers
- Cost-sensitive developers (no paid API required)
- Academic researchers
Competitive Advantages
- Fully open-source and free (MIT license), no usage restrictions
- Best-in-class Chinese reranking performance
- v2-m3 is the de facto standard for open-source multilingual reranking
- Multiple model sizes available, adaptable to different hardware
- v2.5 lightweight version reduces resource consumption
- Seamless integration with BGE embedding models, a one-stop open-source retrieval solution
- Flexible deployment (local/cloud/Azure/NVIDIA NIM)
Limitations
- Higher latency than embedding models due to cross-encoder architecture
- Large models (gemma2-9B base) require good GPU resources
- Lack of commercial-grade API and SLA
- Limited community technical support
- Documentation primarily in English
Relationship with OpenClaw Ecosystem
BGE Reranker is the best open-source choice for local reranking in OpenClaw. bge-reranker-v2-m3 can run entirely on user devices without any API calls, perfectly protecting privacy. Its Chinese reranking capability is crucial for OpenClaw users in China. Combined with the BGE-M3 embedding model, it can build a completely free, fully local embedding + reranking retrieval pipeline, making it an ideal solution for OpenClaw personal use scenarios.
External References
Learn more from these authoritative sources: