kNN search with e5-large data
#78
by
wshan13
- opened
We currently have a system using e5-large for embeddings.
We're considering switching to BGE-M3 due to its better performance.
The embedding vector values of e5-large and BGE-M3 are different,
but I'm wondering if it's possible to perform kNN search on passage data embedded with e5-large using BGE-M3 query embeddings.
While it may not be as accurate as searching data embedded with BGE-M3,
it seems to work to some extent, and I'm curious about the underlying principle.
For example, could it be that although the vector values are different,
they occupy similar topological spaces, or perhaps their dot products are similar, or maybe they were trained on the same data, etc.?