Model Storage

You can load your trained model in three different ways:

1. Load the model from your local disk (see Load Model from Disk)

2. Fetch the model from your own HTTP server (see Load Model from Server)

3. Fetch the model from cloud storage like S3 (see Load Model from Cloud)

By default all commands of Rasa's CLI will load models from your local disk.

Load Model from Disk

By default, models will be loaded from your local disk. You can specify the path to your model with the --model parameter:

rasa run --model models/20190506-100418.tar.gz

If you want to load the latest model in a directory, you can specify a directory instead of a file:

rasa run --model models/

Rasa will check all the models in that directory and load the one that was trained most recently.

If you don't specify a --model argument, Rasa will look for models in the models/ directory. The two following calls will load the same model:

# this command will load the same model
rasa run --model models/
# ... as this command (using defaults)
rasa run

Load Model from Server

You can configure the Rasa server to regularly fetch a model from a server and deploy it.

How to Configure Rasa

You can configure the HTTP server to fetch models from another URL by adding it to your endpoints.yml:

endpoints.yml

models:
  url: http://my-server.com/models/default
  wait_time_between_pulls: 10   # In seconds, optional, default: 100

The server will query the url for a zipped model every wait_time_between_pulls seconds.

If you want to pull the model only when starting up the server, you can set the time between pulls to null:

endpoints.yml

models:
  url: http://my-server.com/models/default
  wait_time_between_pulls: null  # fetches model only once

How to Configure Your Server

Rasa will send a GET request to the URL you specified in the endpoints.yml, e.g. http://my-server.com/models/default in the above examples. You can use any URL. The GET request will contain an If-None-Match header that contains the model hash of the last model it downloaded. An example request from Rasa Open Source to your server would look like this:

curl --header "If-None-Match: d41d8cd98f00b204e9800998ecf8427e" http://my-server.com/models/default

The response of your server to this GET request should be one of these:

• a status code of 200, a zipped Rasa Model and set the ETag header in the response to the hash of the model.
• a status code of 304 and an empty response if the If-None-Match header of the request matches the model you want your server to return.

Rasa uses the If-None-Match and ETag headers for caching. Setting the headers will avoid re-downloading the same model over and over, saving bandwidth and compute resources.

Load Model from Cloud

You can also configure the Rasa server to fetch your model from a remote storage by indicating the remote-storage CLI option when starting the Rasa server. This allows you to retrieve your models from a cloud storage service like Amazon S3, Google Cloud Storage, or Azure Storage.

Ensure you always specify the model name with the --model parameter when running the Rasa server, for example:

rasa run --model 20190506-100418.tar.gz --remote-storage aws

We highly recommend using a dedicated cloud storage bucket for your models.

If your model is stored in a sub folder on your cloud storage, you can specify the path to the model using the --model parameter. Alternatively, you can also provide the environment variable REMOTE_STORAGE_PATH and point to a sub folder within your bucket: note that this latter method is deprecated and will be removed in the next 4.0 major release.

rasa run --model path/to/your/model/20190506-100418.tar.gz --remote-storage aws

The zipped model will be downloaded from cloud storage, unzipped, and deployed. Rasa supports loading models from:

• Amazon S3,
• Google Cloud Storage,
• Azure Storage and
• custom implementations for Other Remote Storages.

Amazon S3 Storage

Amazon S3 is supported using the boto3 package which you need to install as an additional dependency using pip3:

pip3 install boto3

For Rasa to be able to authenticate and download the model, you need to set the following environment variables before running any command requiring the storage:

• AWS_SECRET_ACCESS_KEY: environment variable containing your AWS S3 secret access key

• AWS_ACCESS_KEY_ID: environment variable containing your AWS S3 access key ID

• AWS_DEFAULT_REGION: environment variable specifying the region of your AWS S3 bucket

• BUCKET_NAME: environment variable specifying the S3 bucket

• AWS_ENDPOINT_URL: The complete URL to use for the AWS S3 requests. You need to specify a complete URL (including the "http/https" scheme). For example, you could use one of the Amazon S3 endpoints documented for the region you are using. If you are using a custom endpoint, you can specify it here. Note that by setting the bucket name to BUCKET_NAME environment variable, you should not provide the bucket or object URL to AWS_ENDPOINT_URL.

Once all environment variables are set, you can start the Rasa server with remote-storage option set to aws:

rasa run --model 20190506-100418.tar.gz --remote-storage aws

Google Cloud Storage

Google Cloud Storage (GCS) is supported using the google-cloud-storage package which is already includes in Rasa.

If you are running Rasa on Google App Engine or Compute Engine, the auth credentials are already set up (for the GCS in the same project). In this case, you can skip setting any additional environment variables.

If you are running locally or on a machine outside of GAE or GCE you need to provide the authentication details to Rasa manually:

1. Check out the GCS documentation to create a service account key.
2. Set an environment variable called GOOGLE_APPLICATION_CREDENTIALS to the path of a service account key file with access to your GCS.
3. Set an environment variable called BUCKET_NAME to the name of your GCS bucket.

Once all environment variable is set, you can start the Rasa server with remote-storage option set to gcs:

rasa run --model 20190506-100418.tar.gz --remote-storage gcs

Azure Storage

Azure Storage is supported using the azure-storage-blob package which you need to install as an additional dependency using pip3:

pip3 install azure-storage-blob

For Rasa to be able to authenticate and download the model, you need to set the following environment variables before running any command requiring the storage:

• AZURE_CONTAINER: environment variable containing your azure container name

• AZURE_ACCOUNT_NAME: environment variable containing your azure account name

• AZURE_ACCOUNT_KEY: environment variable containing your account key

Once all environment variables are set, you can start the Rasa server with remote-storage option set to azure:

rasa run --model 20190506-100418.tar.gz --remote-storage azure

Other Remote Storages

If you want to use any other Cloud Storage, you can provide your own python implementation of the rasa.core.persistor.Persistor class, which must inherit from the Persistor interface. This allows you to implement your own logic for retrieving and persisting models to your remote storage.

We recommend you add your custom persistor module to the root of the assistant project before starting the Rasa server, so that Rasa can find your custom persistor class. For example, this is how your project structure could look like:

my_assistant_project/
├── actions/
│   ├── __init__.py
│   ├── actions.py
├── data/
├── models/
├── tests/
├── domain.yml
├── config.yml
├── credentials.yml
├── endpoints.yml
├── remote_storage_persistor.py  # Your custom persistor module

You can start the Rasa server with remote-storage option set to the module path of your persistor implementation:

rasa run --remote-storage <your module>.<class name> --model 20190506-100418.tar.gz

If you are running the Rasa server with Docker, you can mount the custom persistor module path to the container using the -v option:

docker run -v path/to/your/assistant/folder:/app --env_file=path/to/your/env/file rasa/rasa-pro:<RASA_VERSION> run --remote-storage <your module>.<class name> --model 20190506-100418.tar.gz

Persistor Interface

The Persistor interface requires you to implement the following methods:

import abc

from typing import Optional


class Persistor(abc.ABC):
    @abc.abstractmethod
    def _retrieve_tar(self, filename: str, target_path: Optional[str] = None) -> None:
        """Downloads a model previously persisted to cloud storage.

        Args:
            filename: The name of the file in cloud storage.
            target_path: The path on disk where the model should be downloaded to.
        """
        raise NotImplementedError

    @abc.abstractmethod
    def _persist_tar(self, file_key: str, tar_name: str) -> None:
        """Uploads a model to cloud storage.

        Args:
            file_key: The file path to be persisted in cloud storage.
            tar_name: The name of the model that should be uploaded.
        """
        raise NotImplementedError

    @abc.abstractmethod
    def _retrieve_tar_size(self, filename: str, target_path: Optional[str] = None) -> int:
        """Returns the size of the model that has been persisted to cloud storage."""
        raise NotImplementedError

Example Persistor Implementations

Here are some example custom implementations of the Persistor interface for the following remote storage types:

• JFrog Artifactory
• Sonatype Nexus Repository

JFrog Artifactory

You can use the JFrog Artifactory as a remote storage for your models. To use it, you can implement the Persistor interface:

my_custom_persistor.py

from tempfile import NamedTemporaryFile
from typing import Optional

import requests
from rasa.core.persistor import Persistor


class JFrogArtifactoryPersistor(Persistor):
    def _retrieve_tar(self, filename: str, target_path: Optional[str] = None) -> None:
        """Downloads a model previously persisted to cloud storage.

        Args:
            filename: The name of the file in cloud storage.
            target_path: The path on disk where the model should be downloaded to.
        """
        # add auth header for JFrog Artifactory
        headers = {
            "X-JFrog-Art-API": "<replace_with_your_reference_token>"
        }
        response = requests.get(
            f"https://<jfrog_instance_placeholder>/artifactory/<repo_key_placeholder>/{filename}",
            headers=headers,
        )
        if response.status_code != 200:
            raise ValueError(
                f"Failed to retrieve model {filename}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )

        with NamedTemporaryFile(dir=target_path, suffix=".tar.gz", delete=False) as file:
            file.write(response.content)
            file.flush()

    def _persist_tar(self, file_key: str, tar_name: str) -> None:
        """Uploads a model to cloud storage.

        Args:
            file_key: The file path to be persisted in cloud storage.
            tar_name: The name of the model that should be uploaded.
        """
        # add auth header for JFrog Artifactory
        headers = {
            "X-JFrog-Art-API": "<replace_with_your_reference_token>"
        }
        with open(tar_name, "rb") as file:
            data = file.read()
            response = requests.put(
                f"https://<jfrog_instance_name_placeholder>/artifactory/<repo_key_placeholder>/{file_key}",
                headers=headers,
                data=data,
            )
        if response.status_code != 201:
            raise ValueError(
                f"Failed to persist model {tar_name}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )

    def _retrieve_tar_size(self, filename: str, target_path: Optional[str] = None) -> int:
        """Returns the size of the model that has been persisted to cloud storage."""
        # add auth header for JFrog Artifactory
        headers = {
            "X-JFrog-Art-API": "<replace_with_your_reference_token>"
        }
        response = requests.head(
            f"https://<jfrog_instance_name_placeholder>/artifactory/<repo_key_placeholder>/{filename}",
            headers=headers,
        )
        if response.status_code != 200:
            raise ValueError(
                f"Failed to retrieve model size for {filename}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )
        return int(response.headers.get("Content-Length", 0))

Then run the Rasa server with the remote-storage option set to your custom persistor class. For example, if you have implemented the above custom persistor class called JFrogArtifactoryPersistor in a module my_custom_persistor.py, you can run the Rasa server like this:

rasa run --remote-storage my_custom_persistor.JFrogArtifactoryPersistor --model path/to/<model_name>.tar.gz

Sonatype Nexus Repository

You can use the Sonatype Nexus Repository as a remote storage for your models. To use it, you can implement the Persistor interface:

my_custom_persistor.py

import json
from pathlib import Path
from tempfile import NamedTemporaryFile
from typing import Optional

import requests
from rasa.core.persistor import Persistor


REPOSITORY = "<repository_name_placeholder>"  # Replace with your Nexus repository name
AUTH = ("username", "password")  # Replace with your Nexus credentials
NEXUS_DOMAIN = "localhost:8081"  # Replace with your Nexus domain

class SonatypeNexusRepositoryPersistor(Persistor):
    def _retrieve_tar(self, filename: str, target_path: Optional[str] = None) -> None:
        response = requests.get(
            f"http://{NEXUS_DOMAIN}/repository/{REPOSITORY}/{filename}",
            auth=AUTH,
        )
        print(f"Downloading {filename}...")

        if response.status_code != 200:
            raise ValueError(
                f"Failed to retrieve model {filename}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )

        with NamedTemporaryFile(dir=target_path, suffix=".tar.gz", delete=False) as file:
            file.write(response.content)
            file.flush()

    def _persist_tar(self, file_key: str, tar_name: str) -> None:
        with open(tar_name, "rb") as file:
            data = file.read()
            response = requests.put(
                f"http://{NEXUS_DOMAIN}/repository/{REPOSITORY}/{tar_name}",
                auth=AUTH,
                data=data,
            )
        if response.status_code != 201:
            raise ValueError(
                f"Failed to persist model {tar_name}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )

    def _retrieve_tar_size(self, filename: str, target_path: Optional[str] = None) -> int:
        response = requests.head(
            f"http://{NEXUS_DOMAIN}/repository/{REPOSITORY}/{filename}",
            auth=AUTH
        )
        if response.status_code != 200:
            raise ValueError(
                f"Failed to retrieve model size for {filename}. "
                f"Status code: {response.status_code}, Response: {response.text}"
            )
        return int(response.headers.get("Content-Length", 0))

Then run the Rasa server with the remote-storage option set to your custom persistor class. For example, if you have implemented the above custom persistor class called SonatypeNexusRepositoryPersistor in a module my_custom_persistor.py, you can run the Rasa server like this:

rasa run --remote-storage my_custom_persistor.SonatypeNexusRepositoryPersistor --model path/to/<model_name>.tar.gz

Save Model To Cloud

New in 3.10

You can now also save models to cloud after training.

You can also configure the Rasa server to upload your model to a remote storage:

rasa train --fixed-model-name 20190506-100418.tar.gz --remote-storage aws

We highly recommend using a dedicated cloud storage bucket for your models.

note

It is useful to provide a fixed model name while pushing model to remote storage, as you can refer the same name while downloading and running the rasa bot. If no fixed model name is provided, rasa will generate a model name and upload it to remote storage.

Rasa supports uploading models to:

• Amazon S3,
• Google Cloud Storage,
• Azure Storage and
• custom implementations for Other Remote Storages.

You can specify the path on the cloud storage by using the --out parameter during training, along with the --remote-storage option:

rasa train --fixed-model-name 20190506-100418.tar.gz --remote-storage aws --out path/to/models

If the --out parameter is not specified, the model will be uploaded to the default /models directory in the cloud storage bucket.