用 Jev 做检索排序而非 embedding 的建议

Shardul @ShardulAggarwal原推发布于 2026.09.26

在 X 上打开原推

There's a reason we use higher dimensional embeddings trained on search related tasks. you are losing all the information from these strings and the results are gonna be less accurate. Jev wasn't trained to produce embeddings that are accurate for downstream search. use any bi-encoder search embedding models for retrieval and then rank the top-k documents using jev. This is gonna be much more accurate. Just using embeddings is already gonna be much cheaper and more accurate than what you proposed.