Query Instructions for bge-small-en-v1.5

#4
by kazmi09 - opened

Hi,
I am using bge-small-en-v1.5 embedding model for semantic search use case. I am bit confuse about the usage of query instructions as in the model card it is recommended to add "Represent this sentence for searching relevant passages:" prefix in user query. So, should we always add the same prefix against each user query? or we are supposed to add dynamic prefixes depending on the user query?

@Shitao It would be great if you can respond to this thread.

Beijing Academy of Artificial Intelligence org

Hi, just adding the same prefix against each user query is okay.

okay great, Thanks for confirming @Shitao . just a side note, we have also used the old version i-e bge-small-en and it gives more promising results than the updated model.
As mentioned in the documentation the version 1.5 performs better than the older one, but in our case bge-small-en is giving more relevant results.

Are we missing something, or it depends on the dataset we are using?

Beijing Academy of Artificial Intelligence org

we improve the similarity distribution of version 1.5, but the retrieval performance may not be better than version 1.0. You can select the model which performers better on your data.

Sign up or log in to comment