Thanks to IBM just released a very light weight model and the internet offered great instruction for tuning.
I tried with some custom support agent training data specifically related to my blog and wordpress, and embeded the model into a widget. Inference is very slow but testable. I squeezed everything Nexa-Vector 9, Nexa-Vector 10 and this (I call it Z-bot) into a micro machine (budget constraint :(, also there seems some more work I need to do :(.
update Oct 15th seems working better now 
Leave a Reply