it looks like a really simple embed query function but i think this is the key to major improvements
like , what if i could improve the query by adding extra words in that query
basically think like a query router , you see what kind of query it is and add certain more words/tokens to the already existing query to better the embeddings and have better search and stuff.
i can do some experiments and add some kind of tests etc to see its score and how much it is.
drop this implementation of 'query router for token addition' incase i dont see SIGNIFICANT better responses.
LAST MOST IMPORTANT THING
try to keep the token addition sensitive to miniLM L6 , see what the embedding model behaves the best with
basically i think this would turn out good if i actually run it and then log down the behaviour
id have to manually see the likeliness of what i wanted in the query +tests.
so whatver i do here int he query router is slightly embedding model specific , but not too much since they have a good corpus in english.
it looks like a really simple embed query function but i think this is the key to major improvements
like , what if i could improve the query by adding extra words in that query
basically think like a query router , you see what kind of query it is and add certain more words/tokens to the already existing query to better the embeddings and have better search and stuff.
i can do some experiments and add some kind of tests etc to see its score and how much it is.
drop this implementation of 'query router for token addition' incase i dont see SIGNIFICANT better responses.
LAST MOST IMPORTANT THING
try to keep the token addition sensitive to miniLM L6 , see what the embedding model behaves the best with
basically i think this would turn out good if i actually run it and then log down the behaviour
id have to manually see the likeliness of what i wanted in the query +tests.
so whatver i do here int he query router is slightly embedding model specific , but not too much since they have a good corpus in english.