search

netease-youdao / QAnything

Question and Answer based on Anything.
https://qanything.ai
GNU Affero General Public License v3.0
11.93k stars 1.16k forks source link

2.0里分数筛选这里是不是反了？ #528

Open hlqzc2008 opened 2 months ago

hlqzc2008 commented 2 months ago

https://github.com/netease-youdao/QAnything/blob/62c7019b1d11c3c6bde5a674309ae6482508a1a3/qanything_kernel/core/local_doc_qa.py#L550

亲测相近文本的分数更低，这里是否应该是“<= 0.9”？

hlqzc2008 commented 2 months ago

https://github.com/netease-youdao/QAnything/blob/62c7019b1d11c3c6bde5a674309ae6482508a1a3/qanything_kernel/core/retriever/parent_retriever.py#L217

使用search_type=“mmr”时分数是越高越相近，是否应该使用“mmr”而非“similarity”？