Hello rasa community,
I’m new to rasa and setup an environment :
rasa 2.3.0
rasa-sdk 2.3.1
rasa-x 0.35.1
to have a quick test for a
Usecase:
"We have a lot of different printed/pdf forms for different applications, requests etc. like in a government agency. To help people to find the right form for their request i would use rasa”
Approach:
I created for each form an intent and generated typical questions for this intend with keywords and synonyms about the forms topic.
(Some more input were taken from the division responsible for the form)
At the end I got ~350 intents (a lot of forms we use) with around 200 typical question for 1 form → 68.000 lines in the nlu.yml
to keep it short i just created rules (simple intent,action with a corresponding utter_msg with link to the form) for each form, no stories.
Training was fine 15h-> 1.5h after I got the GPU working
Now to the questions / problems: 
Questions :
- Is this a valid approach or would you suggest something different ?
- Should i use synonyms instead of creating lines in the nlu.yml (save training time ? (didn’t try) use less power , whatever or is it rasa-internal the same thing ? )
- are 350 intents with 200 lines training to much ? (see problem)
Problem:
- starting rasa x with the model takes ages, ~4h. log show’s: even 1 step insection(hope that’s the right word ) of the 350 intents took around 1.5 h
Thank you Alex.