cross-encoder/ms-marco-MiniLM-L-6-v2

A high-speed re-ranking model optimized for passage retrieval using the MS MARCO dataset.

This model utilizes a Cross-Encoder architecture to score the relevance of document-query pairs with significantly higher accuracy than standard Bi-Encoder embeddings. Built on the MiniLM-L-6-v2 backbone, it strikes a precise balance between performance and latency, processing inputs through 6 transformer layers. It is specifically designed as a second-stage re-ranker in retrieval pipelines (like Elasticsearch or Pinecone) to refine the top-k results returned by initial semantic searches.

https://huggingface.co/cross-encoder/ms-marco-MiniLM-L-6-v2