Currently struggling with an automatic budget selection, at 25% it’s very similar to benchmarks accuracy and seems even better on actual chat input however it many times retrieves too much. It would be nice to add an algorithm that actually can determine better retrieval other then CELF.
