A Java based Rasa Action Server

The rasa-java-action-service is an environment for building your Rasa Custom Actions with Java.

• Build your Rasa assistant fulfillment easily: your Actions just implement a simple Java interface
• With full Rasa stack onboard: set up Rasa and connect to your Action Server with only one command
• Control the Rasa CLI with simple HTTP API calls

The rasa-java-action-service is inspired by Rafał Bajek's Action Server implementation and uses his SDK

current version: 0.2.0 (unstable)

The rasa-java-action-service is not a 'ready to use' application. It functions as a template and starting point to build your Rasa assistant's fulfillment. Therefore, just clone this repo and start.

You need to install Docker and Docker Compose.

Use the rasa-java-action-service if you consider building your Rasa Action Server with Java.

Almost every advanced, business integrated conversational AI system has the ability to perform business logic procedures during a conversation with the user - be it a calculation, a database query or an API call. This ability is called fulfillment and is typically organized in so-called Actions. An Action is a part of code which is responsible for one independent business logic procedure.

In Rasa, Actions are placed within the Rasa Action Server. If Rasa predicts an Action to be executed, it calls the Action Server which executes the Action thereupon and responds with an enriched response object to the Rasa Server.

The rasa-java-action-service provides an initial set-up configuration which makes a quick development start possible.

• clone this repo
• go to project directory
• in the command line type docker-compose -f docker-compose.develop.yml up -d
  • after an image build this will run a docker container named rasa_server_dev starting the Rasa Server with API enabled listening on localhost:5005
  • the rasabot folder is mounted in the container's /app directory
• the rasabot folder contains an initial set of files which you can use as a starting point to build your Rasa assistant
• type now docker exec -it rasa_server_dev bash to get the container shell
• type rasa train into the container shell
  • this will train a Rasa model based on the content of the .yml files in the rasabot folder
  • the .yml files are initially prefilled with content which defines a sample Rasa bot
  • the sample bot is an extended Rasa 'mood bot' which asks you how you are, and tries to cheer you up if you feel bad.
• start the RasaActionServer (within IDE or after a build withjava -jar). The Action Server listens per default on port 5055, see application.properties
  • initially the RasaActionServer contains an sample Action named ActionSample which retrieves a joke though an API call.
  • Rasa tries to predict this Action and calls the Action Server if you tell the bot that you feel bad
• after training completed, type rasa shell -p 5006 to load the model. After loading start the conversation by greeting the bot. The bot will ask you, how you are.
• if you tell the bot you feel sad, the bot will try to cheer you up with a joke and call the Action Server.
• if the bot responds with a joke, then the Action Server connection works, and your set-up was successful.

See the Rasa Docs if you want to know how to build conversational assistants with Rasa.

Build your Actions with the rasa-java-action-service by just implementing the Action interface and declaring them with the Spring @Component annotation:

@Component
public class ActionSample implements Action {
    
    @Override
    public String name() {
        return "<name of you Action as defined in domain.yml file>";
    }

    @Override
    public List<AbstractEvent> run(CollectingDispatcher dispatcher, Tracker tracker, Domain domain) {
        
        List<AbstractEvent> eventList = null;
        //your Custom Action code
        //...
        
        //provide your responses
        dispatcher.utterMessage("<your message>");
        
        //provide events
        return eventList;
    }

}

See also the sample Action ActionSample in the actions package.

For local development just edit the .yml files in the rasabot folder. Afterwards you need to re-train your bot. See start section how to train the bot and start a conversation in the Rasa Shell.

If your Actions are implemented, then your Action Server is ready for deployment. You can deploy your Action Server everywhere where you have Docker and Docker Compose installed and access to your project.

• uncomment the http://actionservice:5055/webhook url in action_endpoint section of the endpoints.yml file. Comment the other urls in the action_endpoint section.
• go to the project directory
• type in the command line: docker-compose up -d. This will result in the following steps:
  • a multi-stage image build for the Action Server will be triggered and the action_server container will be started listening on exposed port 5055
  • a multi-stage image build for an extended Rasa environment will be triggered and the rasa_server container will be started listening on exposed port 5005
  • within the rasa_server container a flask application will be started listening on exposed port 4000. Attention! This is an experimental feature! It is used to access the Rasa CLI through an API endpoint.
  • action_server and rasa_server are connected with a shared named volume. You can see the same content in the /app directory of rasa_server as well as in the /app/rasabot directory of action_server
    • the shared volume is used because Rasa doesn't provide any API endpoints to manipulate the endpoints.yml and the credentials.yml files. With the shared volume configuration your Action Server can be extended with corresponding API endpoints to solve this problem.

After your Action Server and Rasa are deployed, you can talk with or train your Rasa bot either in the Rasa Shell as shown in the start section or using the Rasa Open Source HTTP API.

If you want to build your Rasa assistant with Java you should use the rasa-java-client-library.

• Use the comfortable RasaClient to connect to the Rasa Server and talk to or train your Rasa assistant
• Use the flexible ModelApi, DomainApi and TrackerApi to implement additional advanced features, e.g. interactive learning

Within the rasa_server container a flask application will be started listening on exposed port 4000. Attention! This is an experimental feature! The flask server provides an endpoint called http://localhost:4000/commands/rasa and can be used to control Rasa by requesting Rasa CLI commands. Interactive commands like rasa shell or rasa interactive won't work.

A request like this:

curl -X POST -H 'Content-Type: application/json' -d '{"args": ["train", "--num-threads", "4"]}' http://localhost:4000/commands/rasa

... will return a JSON response like this:

{
   "key": "123456",
   "result_url": "http://localhost:4000/commands/rasa?key=123456",
   "status": "running"
}

now you can poll on the delivered key:

curl http://localhost:4000/commands/rasa?key=123456

... to get the result when the command returns:

{
  "report": "<rasa train output>",
  "key": "123456",
  "start_time": 1593019807.7754705,
  "end_time": 7593019807.782958,
  "process_time": 6.00748753547668457,
  "returncode": 0,
  "error": null
}