Ability to compose domain files

[moaazsidat]

Description of Problem: Currently there's no opinions that Rasa enforces about being able to split the domain file into multiple files (unlike stories e.g.).

We ran into this issue when trying to model a rule based system within a chatbot where we ended up with 3000+ slots and 1000+ actions for that functionality within the chatbot. This main the domain file increasingly frustrating to add to and work with.

Overview of the Solution: Spent some time last night writing a domain composer for our use case and wanted to get thoughts from the community about it.

here's the structure of my data/ folder:

├── domain.yml
├── domains
│   ├── hotel_booking
│   │   ├── actions.yml
│   │   ├── forms.yml
│   │   ├── intents.yml
│   │   └── slots.yml
│   ├── common
│   │   ├── actions.yml
│   │   ├── forms.yml
│   │   ├── intents.yml
│   │   └── slots.yml
│   ├── flight_booking
│   │   ├── actions.yml
│   │   ├── forms.yml
│   │   ├── intents.yml
│   │   └── slots.yml
└── stories.md

when a dev runs make build:domain which uses the composer script aforementioned they get a data/domain.yml file, which can then be used for training, that looks like this:

actions:
  <all action values>
forms:
  <all form values>
intents:
  <all intent values>
slots:
  <all slot values>

Questions

• Wondering if this should be natively supported within the rasa training toolchain similar to how splitting up stories into multiple files is supported
• What are the limitations of doing this?

[sara-tagger]

Thanks for submitting this feature request 🚀@MetcalfeTom will get back to you about it soon!✨

[MetcalfeTom]

Hi @moaazsidat, like the idea :+1:

How exactly does the training data importing work here? Does your stories.md contain stories from all three of the domains in your example?

Seems like your solution deals implements two features here:

1. I like the domain-splitting for larger projects. Seems like an intuitive way of making large projects digestible.

2. The other feature you're proposing is the unification of a multi-skill assistant. This is already in the code, but is still experimental

Unfortunately we merged the ability to make custom Training Data Importers with rasa==1.1.8 and I think your implementation is already covered by this functionality. Give it a test and let us know if you have any feedback 😄I wonder if you could instead be searching for these domain files and unifying them automatically by passing a domain directory, rather than a make command.

[no-response[bot]]

This issue has been automatically closed because there has been no response to our request for more information from the original author. Without this, we don't have enough information to help you. Please comment below with the requested information if you still need help.