BGE-Reranker

Different from embedding model, reranker, or cross-encoder uses question and document as input and directly output similarity instead of embedding. To balance the accuracy and time cost, cross-encoder is widely used to re-rank top-k documents retrieved by other simple models. For examples, use a bge embedding model to first retrieve top 100 relevant documents, and then use bge reranker to re-rank the top 100 document to get the final top-3 results.

The first series of BGE-Reranker contains two models, large and base.

Model	Language	Parameters	Model Size	Description
BAAI/bge-reranker-large	English & Chinese	560M	2.24 GB	Larger reranker model, easy to deploy with better inference
BAAI/bge-reranker-base	English & Chinese	278M	1.11 GB	Lightweight reranker model, easy to deploy with fast inference

bge-reranker-large and bge-reranker-base used XLM-RoBERTa-Large and XLM-RoBERTa-Base respectively as the base model. They were trained on high quality English and Chinese data, and acheived State-of-The-Art performance in the level of same size models at the time released.

Usage

from FlagEmbedding import FlagReranker

reranker = FlagReranker(
    'BAAI/bge-reranker-base',
    query_max_length=256,
    use_fp16=True,
    devices=['cuda:1'],
)

score = reranker.compute_score(['I am happy to help', 'Assisting you is my pleasure'])
print(score)