Tag
Single-vector embedding models can be used to extract sparse latent terms, and BM25 can turn this vocabulary into a strong retriever.