Skip to content

rasa.md

Latest commit

 

History

History
81 lines (56 loc) · 4.26 KB

rasa.md

File metadata and controls

81 lines (56 loc) · 4.26 KB

Rasa

IMPORTANT. Install Python 3.9 first as Rasa does not support later versions yet (as of July 2022).

Automatic launch

You can start Rasa by calling python3 rasa_launcher.py from the root directory.

Manual launch

Go to "rasa" folder, create virtual environment and switch to the created virtual environment:

cd rasa
python3.9 -m venv ./venv
source ./venv/bin/activate

Install latest version of pip: pip3 install -U pip

Install Rasa by running pip3 install rasa or install dependencies from requirements.txt by running pip3 install -r requirements.txt.

Install spaCy (currently, spaCy model is used by the replica's pipeline):

pip3 install rasa[spacy]
python3 -m spacy download en_core_web_lg

Go to the bot folder. It is a Rasa project folder. It contains data for the bot. Run all the Rasa-related commands from this folder.

Use rasa train --domain domain to train your model. The trained model will appear in the models folder.

Start Rasa server: rasa run -i localhost -p 8002.

As per default config, the AI Replica server expects the Rasa server is being run at http://localhost:8002. By default, the latest model from the models folder is used. If you want to run a specific model, you can do it as follows: rasa run -i localhost -p 8002 --model models/20220122-164843-charitable-rent.tar.gz.

Start Rasa actions server as follows: SANIC_HOST=localhost rasa run actions -p 8004.

Debug Rasa

You can get access to some useful Rasa internal info, e.g. intent confidence, via Rasa http api. To activate the api, start Rasa server with --enable-api flag, e.g. rasa run -i localhost -p 8002 --enable-api.

Then you can, for example, access conversation events via calling http://localhost:8002/conversations/user/tracker. Note. Here, 'user' is the value of the 'sender' field sent by bot server to Rasa webhook api (see request_handler.py).

Check Tracker api here: https://rasa.com/docs/rasa/pages/http-api/#tag/Tracker.

Also, you can run Rasa with --debug flag to see some useful output in the console. E.g. rasa run -i localhost -p 8002 --debug.

Potentially, it is even more convenient to install Rasa X for the purpose of debugging but it is intended primarily for server-mode setup, so the installation and running is not so straightforward.

Notes

  1. Command line interface: https://rasa.com/docs/rasa/command-line-interface
  2. Installation: https://rasa.com/docs/rasa/installation/. Install Python 3.8 first as Rasa does not support later versions yet (as of January 2022). In the pip3 install -U --user pip && pip3 install rasa command remove --user flag, otherwise you can experience an error saying that execution is not possible ("Can not perform a '--user' install. User site-packages are not visible in this virtualenv."). There is no sense to install packages to user folder (--user) when you are in a virtual environment.

Useful stuff

rasa shell command starts bot in a terminal. --debug flag passed to a rasa command allows seeing some debug information.

Conversation stores

Sqlite

Install (on Ubuntu): sudo apt install sqlite3 Create a folder where you db will be stored. sqlite does not create folder and throws an error if the folder does not exist. The db file is created by sqlite itself. Configure: Tracker Stores

tracker_store:
   type: SQL
   dialect: sqlite
   url: sqlite:///../../../db/rasa.db # specify path to a db file after sqlite:///
   db: rasa.db
   username: 
   password:

API

webhook

By default, Rasa server provides webhook api: see, see. It looks like webhook is a conversation channel, i.e. it allows to send input messages from user and receive replies from the bot but does not provide access to anything else. E.g. you can have the following webhooks (channels): rest, socket.io, slack, etc.

http api

To have access to additional functionality of a Rasa server, you can start server with exposed api: Http API, Note Make sure that the api endoints are protecte either with authorization or hidden behind a proxy server/app router, so that they are not exposed to the Internet.