bge 模型深入使用后的思考与疑问

I am a deep user of bge, and have completed the deployment, inference, fine-tuning, and evaluation of embedding and reranker. Currently, I have discovered a problem and would like to share it with everyone for consideration. First, let me give you my data conclusions, where the key data has been de-identified, and an arrow represents a 5% increase or decrease. An increase or decrease refers to a comparison with the “only embedding, no reranker” scenario: