Reranking is an approach to improve search and recommendation with vector similarity on embeddings.

First, the typical K-nearest neighbours based on vector similiarity are returned. Then, this list is reranked. The reranker ususally used self-attention to find out more details on the relationship of query and knowledge-base.

https://github.com/NirDiamant/RAG_Techniques/blob/main/all_rag_techniques/reranking_with_llamaindex.ipynb

Another explainer on it: https://www.pinecone.io/learn/series/rag/rerankers/#Power-of-Rerankers

Reranking services

Various AI / LLM providers offer

Among them:

• Jina AI
• Cohere
• Qwen