Open eskibars opened 11 months ago
When we experiment with various lambda values, we can try searching for “textless”, “custom dimensions”, and “query” to see how the quality of the results changes.
@cjcenizal yes, we can test with these, but there are even more extreme examples of things that are even rarer. e.g. searching for lxml
or epub
should find https://docs.vectara.com/docs/api-reference/indexing-apis/file-upload/file-upload-filetypes, 272725718
should find the MMR reranker. AdminService
should find https://docs.vectara.com/docs/api-reference/admin-apis/admin
Idea
Technical documentation in particular tends to have a lot of terms that may have never shown up in the neural net training data. This means that not only having a non-zero value for lambda is important, but that actually it needs to be higher than average typically for the best effect. I recommend we start with a value of 0.1 for now
cc @pwoznic -- we also should actually improve the documentation around lambda. Right now it's bundled into hybrid and it's not clear how to really use it from the docs without tab-switching to the playground