BAAI/bge-reranker-v2-m3 · Hugging Face

Overview

BGE Reranker v2 M3 is a cross-encoder reranking model developed by the Beijing Academy of Artificial Intelligence (BAAI). It is part of the BGE (BAAI General Embedding) family and is specifically designed for reranking tasks in information retrieval. The model takes a query and a document pair as input and outputs a relevance score, enabling more accurate ranking of search results. Built on the XLM-RoBERTa architecture, it supports over 100 languages, including low-resource languages, making it suitable for global applications. With over 16 million downloads on Hugging Face, it is widely adopted for multilingual and cross-lingual retrieval scenarios. The model is associated with two papers: "Making Large Language Models A Better Foundation For Dense Retrieval" (arxiv:2312.15503) and "BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation" (arxiv:2402.03216).

Capabilities

• Multilingual Support: Supports 100+ languages, including low-resource languages, making it suitable for global applications. The model is tagged as multilingual on Hugging Face.
• Cross-Encoder Architecture: Provides higher accuracy than bi-encoder models by jointly encoding query and document, allowing for deeper interaction between the two inputs.
• State-of-the-Art Performance: Achieves top results on benchmarks like BEIR, MIRACL, and multilingual retrieval tasks. Specific benchmark scores are not detailed in the source, but the model is noted for competitive performance.
• Efficient Inference: Optimized for text-embeddings-inference and compatible with Hugging Face pipelines. It can be deployed on CPU or GPU, with the model weights stored in safetensors format (567 million parameters, 2.27 GB).
• License: Apache 2.0, allowing commercial use, modification, and distribution.
• Integration: Available via Hugging Face Transformers, Sentence-Transformers, and Text Embeddings Inference. It is also compatible with Hugging Face Inference Endpoints and Azure deployment.

Use cases

• Enterprise Search: Improve search relevance in multilingual knowledge bases and intranets by reranking initial retrieval results.
• E-commerce: Rerank product search results across different languages to enhance user experience and conversion rates.
• Legal & Compliance: Retrieve relevant documents from multilingual legal databases, ensuring accurate cross-lingual search.
• Academic Research: Enhance cross-lingual information retrieval for scientific papers, enabling researchers to find relevant literature in multiple languages.
• Chatbots & RAG: Rerank retrieved passages in retrieval-augmented generation pipelines, improving the quality of context provided to large language models.

License & deployment

• License: Apache 2.0. This permissive license allows for commercial use, modification, and redistribution, with no restrictions on usage.
• Deployment: Available via Hugging Face Transformers, Sentence-Transformers, and Text Embeddings Inference. Can be deployed on CPU or GPU. The model is compatible with Hugging Face Inference Endpoints and Azure. The official repository is at https://huggingface.co/BAAI/bge-reranker-v2-m3.
• Hardware Considerations: The model has 567 million parameters and requires approximately 2.27 GB of storage. Inference can be performed on CPU, but GPU acceleration is recommended for lower latency. The model is not quantized by default, but quantization options may be available.
• Safety Notes: As a reranking model, it does not generate content but scores relevance. However, biases in training data may affect scores. Users should evaluate fairness and robustness for their specific use case.

Alternatives

• Cohere Rerank: Proprietary reranking API with multilingual support. Offers ease of use but is not open-source and requires API costs.
• Cross-Encoder/ms-marco-MiniLM-L-6-v2: Lightweight English-only reranker based on MiniLM. Suitable for English-only applications with lower resource requirements.
• BGE Reranker v2.0: Larger variant from BAAI with higher accuracy but more resource-intensive. May be preferred when maximum performance is needed.
• Jina Reranker: Open-source multilingual reranker with competitive performance. Offers similar capabilities but may differ in language coverage and efficiency.

FAQ

Q: What is the difference between BGE Reranker v2 M3 and other BGE models? A: BGE Reranker v2 M3 is a cross-encoder for reranking, while other BGE models (e.g., BGE Embedding) are bi-encoders for embedding generation. The M3 variant is optimized for multilingual and cross-lingual retrieval, supporting 100+ languages.

Q: Can this model be used for embedding generation? A: No, this model is specifically a cross-encoder for reranking. For embedding generation, BAAI provides bi-encoder models like BGE Embedding.

Q: What is the input format? A: The model takes a query and a document as a pair, typically formatted as "[CLS] query [SEP] document [SEP]". The output is a relevance score (logit).

Q: Is the model available for commercial use? A: Yes, it is licensed under Apache 2.0, which permits commercial use.

Q: What hardware is recommended? A: The model can run on CPU but GPU is recommended for faster inference. It has 567 million parameters and requires about 2.27 GB of memory.

Q: How does it compare to Cohere Rerank? A: BGE Reranker v2 M3 is open-source and free, while Cohere Rerank is a proprietary API with usage costs. Both support multilingual reranking, but BGE may offer more flexibility for self-hosting.