Warning: This document is for an old version of Rasa. The latest version is 1.10.1.

Training Data Importers

By default, you can use command line arguments to specify where Rasa should look for training data on your disk. Rasa then loads any potential training files and uses them to train your assistant.

If needed, you can also customize how Rasa imports training data. Potential use cases for this might be:

  • using a custom parser to load training data in other formats

  • using different approaches to collect training data (e.g. loading them from different resources)

You can instruct Rasa to load and use your custom importer by adding the section importers to the Rasa configuration file and specifying the importer with its full class path:

- name: "module.CustomImporter"
  parameter1: "value"
  parameter2: "value2"
- name: "module.AnotherCustomImporter"

The name key is used to determine which importer should be loaded. Any extra parameters are passed as constructor arguments to the loaded importer.


You can specify multiple importers. Rasa will automatically merge their results.

RasaFileImporter (default)

By default Rasa uses the importer RasaFileImporter. If you want to use it on its own, you don’t have to specify anything in your configuration file. If you want to use it together with other importers, add it to your configuration file:

- name: "RasaFileImporter"

MultiProjectImporter (experimental)


This feature is currently experimental and might change or be removed in the future. Please share your feedback on it in the forum to help us making this feature ready for production.

With this importer you can build a contextual AI assistant by combining multiple reusable Rasa projects. You might, for example, handle chitchat with one project and greet your users with another. These projects can be developed in isolation, and then combined at train time to create your assistant.

An example directory structure could look like this:

├── config.yml
└── projects
    ├── GreetBot
    │   ├── data
    │   │   ├── nlu.md
    │   │   └── stories.md
    │   └── domain.yml
    └── ChitchatBot
        ├── config.yml
        ├── data
        │   ├── nlu.md
        │   └── stories.md
        └── domain.yml

In this example the contextual AI assistant imports the ChitchatBot project which in turn imports the GreetBot project. Project imports are defined in the configuration files of each project. To instruct Rasa to use the MultiProjectImporter module, put this section in the config file of your root project:

- name: MultiProjectImporter

Then specify which projects you want to import. In our example, the config.yml in the root project would look like this:

- projects/ChitchatBot

The configuration file of the ChitchatBot in turn references the GreetBot:

- ../GreetBot

The GreetBot project does not specify further projects so the config.yml can be omitted.

Rasa uses relative paths from the referencing configuration file to import projects. These can be anywhere on your file system as long as the file access is permitted.

During the training process Rasa will import all required training files, combine them, and train a unified AI assistant. The merging of the training data happens during runtime, so no additional files with training data are created or visible.


Rasa will use the policy and NLU pipeline configuration of the root project directory during training. Policy or NLU configurations of imported projects will be ignored.


Equal intents, entities, slots, responses, actions and forms will be merged, e.g. if two projects have training data for an intent greet, their training data will be combined.

Writing a Custom Importer

If you are writing a custom importer, this importer has to implement the interface of TrainingDataImporter:

from typing import Optional, Text, Dict, List, Union

import rasa
from rasa.core.domain import Domain
from rasa.core.interpreter import RegexInterpreter, NaturalLanguageInterpreter
from rasa.core.training.structures import StoryGraph
from rasa.importers.importer import TrainingDataImporter
from rasa.nlu.training_data import TrainingData

class MyImporter(TrainingDataImporter):
    """Example implementation of a custom importer component."""

    def __init__(
        config_file: Optional[Text] = None,
        domain_path: Optional[Text] = None,
        training_data_paths: Optional[Union[List[Text], Text]] = None,
        **kwargs: Dict
        """Constructor of your custom file importer.

            config_file: Path to configuration file from command line arguments.
            domain_path: Path to domain file from command line arguments.
            training_data_paths: Path to training files from command line arguments.
            **kwargs: Extra parameters passed through configuration in configuration file.


    async def get_domain(self) -> Domain:
        path_to_domain_file = self._custom_get_domain_file()
        return Domain.load(path_to_domain_file)

    def _custom_get_domain_file(self) -> Text:

    async def get_stories(
        interpreter: "NaturalLanguageInterpreter" = RegexInterpreter(),
        template_variables: Optional[Dict] = None,
        use_e2e: bool = False,
        exclusion_percentage: Optional[int] = None,
    ) -> StoryGraph:
        from rasa.core.training.dsl import StoryFileReader

        path_to_stories = self._custom_get_story_file()
        return await StoryFileReader.read_from_file(path_to_stories, await self.get_domain())

    def _custom_get_story_file(self) -> Text:

    async def get_config(self) -> Dict:
        path_to_config = self._custom_get_config_file()
        return rasa.utils.io.read_config_file(path_to_config)

    def _custom_get_config_file(self) -> Text:

    async def get_nlu_data(self, language: Optional[Text] = "en") -> TrainingData:
        from rasa.nlu.training_data import loading

        path_to_nlu_file = self._custom_get_nlu_file()
        return loading.load_data(path_to_nlu_file)

    def _custom_get_nlu_file(self) -> Text:


class rasa.importers.importer.TrainingDataImporter

Common interface for different mechanisms to load training data.

async get_domain()

Retrieves the domain of the bot.


Loaded Domain.

Return type


async get_config()

Retrieves the configuration that should be used for the training.


The configuration as dictionary.

Return type

Dict[~KT, ~VT]

async get_nlu_data(language='en')

Retrieves the NLU training data that should be used for training.


language – Can be used to only load training data for a certain language.


Loaded NLU TrainingData.

Return type


async get_stories(interpreter=<rasa.core.interpreter.RegexInterpreter object>, template_variables=None, use_e2e=False, exclusion_percentage=None)

Retrieves the stories that should be used for training.

  • interpreter – Interpreter that should be used to parse end to end learning annotations.

  • template_variables – Values of templates that should be replaced while reading the story files.

  • use_e2e – Specifies whether to parse end to end learning annotations.

  • exclusion_percentage – Amount of training data that should be excluded.


StoryGraph containing all loaded stories.

Return type