rasa-project: A Dockerfile repository from RedTint

How to setup

Run pip3 install rasa==2.0.0rc2. NOTE: This install a lot of dependencies, you may want to use virtualenv to keep your local dependencies organized
Run rasa train. This should train the Rasa Assistant, with 2 initial steps: (1) NLU model training and (2) Core model training
Once training is complete, you will receive a notification that specifies the directory and filename of the model that was trained:

Your Rasa model is trained and saved at '/Users/projects/personal/rasa-project/models/20200927-222831.tar.gz'

NOTE: The model name seems to have been derived from the data and time when rasa train was called. The initial training took 26 mins. for me as it included the download time for the base model: https://github.com/PolyAI-LDN/polyai-models/releases/download/v1.0/model.tar.gz

Rasa will ask you whether you'd like to test the trained rasa assistant on your first try.

Succeeding Training Notes

I tried running traiing again after updating the favorite_color intent and it was able to train it in less than 3 mins.
Tried running again without any changes and I got:

Nothing changed. You can use the old model stored at '/Users/projects/personal/rasa-project/models/20200927-230117.tar.gz'.

Testing Trained Rasa Assistant Model

You can run rasa shell to test the latest trained model.

Automated Testing

Something very unique with RASA is the Automated Testing. Run rasa train to try.

Manually Running the Rasa Server

You can manually run the rasa server by running rasa run --enable-api. See: https://rasa.com/docs/rasa/http-api

Rasa should give you the endpoint you should connect to, e.g.:

Starting Rasa server on http://localhost:5005

Here's the link to the API Documentation: https://rasa.com/docs/rasa/pages/http-api

Here's how to connect the Bot into your Own Website

https://rasa.com/docs/rasa/connectors/your-own-website/#restinput

In a nutshell, you can start conversing with the bot through:

http://localhost:5005/webhooks/rest/webhook

Migrating from DialogFlow

https://rasa.com/docs/rasa/migrate-from

Finding a Model that works with Rasa

After a few debugging here and there, I encountered not being able to run rasa train because of a missing model that is required. Upon further research, I found out that you can download it here:

Working with Docker

Issues

Issue with training model for the first time (./models don't have any model content). For some reason, rasa is not able to store models in the ./models folder despite having to successfully train. This doesn't happen when the ./models folder already contains a model to start.
- We might need to train a model and have the stored indefinitely in our ./models folder as our base?
[RESOLVED] Currently having issues with this:

rasa.exceptions.ModelNotFound: No NLU or Core data for unpacked model at: '/var/folders/4m/6prcxkh10s71tzn6frvdy7lm0000gn/T/tmpk5ekiq3q'

It seems like, it's looking for core and nlu default folders which can't be found when simply running rasa train --fixed-model-name ./models/model.tar.gz.

Steps To Do Again

LOCAL ENVIRONMENT

DOCKERIZED ENVIRONMENT

Using the installation process for Rasa Open Source, build a dockerized environment for working with rasa.

Notes

Response when running rasa run without trained models

$ rasa run
No model found. You have three options to provide a model:
1. Configure a model server in the endpoint configuration and provide the configuration via '--endpoints'.
2. Specify a remote storage via '--remote-storage' to load the model from.
3. Train a model before running the server using `rasa train` and use '--model' to provide the model path.
For more information check https://rasa.com/docs/rasa/model-storage.

Running with rasa run with already trained model logs:

Step 11/14 : RUN echo -e "no\n" | rasa train
 ---> Running in fa8a1582ed4d
2020-09-30 07:48:47 WARNING  rasa.shared.utils.validation  - Training data file data/rules.yml doesn't have a 'version' key. Rasa Open Source will read the file as a version '2.0' file. See https://rasa.com/docs/rasa.
Nothing changed. You can use the old model stored at '/usr/src/app/models/20200929-141249.tar.gz'.

Rasa will always load the server with the latest trained model upon running docker-compose up.
- I attempted to remove the last model and re-run docker-compose and found out that it ran training again. This means that rasa probably keeps track of the last trained model somewhere in cache or some file.
If you delete the models stored in folder ./models, while a model is already loaded, it doesn't affect the Rasa Server.
If no model was loaded, you will get something like this:

{
    "config": {
        "store_entities_as_slots": true
    },
    "session_config": {
        "session_expiration_time": 0,
        "carry_over_slots_to_new_session": true
    },
    "intents": [],
    "entities": [],
    "slots": {},
    "responses": {},
    "actions": [],
    "forms": {},
    "e2e_actions": []
}

Can we run model training via API without a previously trained model? **Answer: ** Yes.
Can we run rasa run without an available pre-trained model? Answer: Yes.

RedTint/rasa-project