Version: Latest

Version Migration Guide

This page contains information about changes between major versions and how you can migrate from one version to another.

Rasa Pro 3.9 to Rasa Pro 3.10

LLM/Embedding Configuration

The LLM and embedding configurations have been updated to use the provider key instead of the type key. These changes apply to all providers, with some examples provided for reference.

Cohere

llm:
  provider: "cohere" # instead of "type: cohere"
  model: "command-r"

Vertex AI

llm:
  provider: "vertex_ai" # instead of "type: vertexai"
  model: "gemini-pro"

Hugging Face Hub

llm:
  provider: "huggingface" # instead of "type: huggingface_hub"
  model: "HuggingFaceH4/zephyr-7b-beta" # instead of "repo_id: HuggingFaceH4/zephyr-7b-beta"

llama.cpp

The support for loading models directly have been removed. You need to deploy the model to a server and use the server URL to load the model. For instance a llama.cpp server can be run using the following command, ./llama-server -m your_model.gguf --port 8080.

For more information on llama.cpp server, refer to the llama.cpp documentation The assistant can be configured as:

llm:
  provider: "self-hosted" # instead of "type: llamacpp"
  api_base: "http://localhost:8000/v1" # instead of "model_path: "/path/to/model.bin""
  model: "ggml-org/Meta-Llama-3.1-8B-Instruct-Q4_0-GGUF"

vLLM

The model can be deployed and served through vLLM==0.6.0. For instance a vLLM server can be run using the following command, vllm serve your_model

For more information on vLLM server, refer to the vLLM documentation The assistant can be configured as:

llm:
  provider: "self-hosted" # instead of "type: vllm_openai"
  api_base: "http://localhost:8000/v1"
  model: "NousResearch/Meta-Llama-3-8B-Instruct"  # the name of the model you have deployed

note

CALM exclusively utilizes the chat completions endpoint of the model server, so it's essential that the model's tokenizer includes a chat template. Models lacking a chat template will not be compatible with CALM anymore.

Backward compatibility has been maintained for OpenAI and Azure configurations. For all other providers, ensure the use of the provider key and review the configuration against the documentation.

Disabling the cache

For Rasa Pro versions <= 3.9.x, the correct way to disable the cache was:

llm:
    model: ...
    cache: false

Rasa Pro 3.10.0 onwards, this has changed since we rely on LiteLLM to manage caching. To avoid errors, change your configuration to -

llm:
    model: ...
    cache:
       no-cache: true

Custom Components using an LLM

As of Rasa Pro 3.10, the backend for sending LLM and Embedding API requests has undergone a significant change. The previous LangChain version 0.0.329 has been replaced with LiteLLM.

This shift can potentially break custom implementations of components that configure and send API requests to chat completion and embedding endpoints. Specifically, the following components are impacted:

If your project contains custom components based on any of the affected components listed above, you will need to verify and possibly refactor your code to ensure compatibility with LiteLLM.

Changes to `llm_factory`

The llm_factory is used across all components that configure and send API requests to an LLM. Previously, the llm_factory relied on LangChain's mapping to instantiate LangChain clients.

Rasa Pro 3.10 onwards, the llm_factory returns clients that conform to the new LLMClient protocol. This impacts any custom component that was previously relying on LangChain types.

If you have overridden components, such as a command generator, you will need to update your code to handle the new return type of LLMClient. This includes adjusting method calls and ensuring compatibility with the new protocol.

The following method calls will need to be adjusted if you have overridden them:

SingleStepLLMCommandGenerator.invoke_llm
MultiStepLLMCommandGenerator.invoke_llm
ContextualResponseRephraser.rephrase
EnterpriseSearchPolicy.predict_action_probabilities
IntentlessPolicy.generate_answer
LLMBasedRouter.predict_commands

Here’s an example of how to update your code:

Rasa 3.9 - LangChain
Rasa 3.10 - LiteLLM

from rasa.shared.utils.llm import llm_factory
from rasa.shared.providers.llm.llm_client import LLMClient
from rasa.shared.providers.llm.llm_response import LLMResponse

# get the llm client via factory
llm: LLMClient = llm_factory(config, default_config)

# get the llm response synchronously
sync_response: LLMResponse = llm.completion(prompt)  # or llm.completion([prompt_1, prompt_2,..., prompt_n])
sync_completion: str = sync_response.choices[0]

# get the llm response asynchronously
async_response: LLMResponse = await llm.acompletion(prompt)  # or llm.acompletion([prompt_1, prompt_2,..., prompt_n])
async_completion: str = async_response.choices[0]

Changes to `embedder_factory`

The embedder_factory is used across all components that configure and send API requests to an embedding model. Previously, the embedder_factory returned LangChain's embedding clients of Embeddings type.

Rasa Pro 3.10 onwards, the embedder_factory returns clients that conform to the new EmbeddingClient protocol. This change is part of the move to LiteLLM, and it impacts any custom components that were previously relying on LangChain types.

If you have overridden components that rely on instantiating clients with embedder_factory you will need to update your code to handle the new return type of EmbeddingClient. This includes adjusting method calls and ensuring compatibility with the new protocol.

The following method calls will need to be adjusted if you have overridden them:

FlowRetrieval.load
FlowRetrieval.populate
EnterpriseSearchPolicy.load
EnterpriseSearchPolicy.train
IntentlessPolicy.load
Or if you have overridden the IntentlessPolicy.embedder attribute.

Here’s an example of how to update your code:

Rasa 3.9 - LangChain
Rasa 3.10 - LiteLLM

from rasa.shared.utils.llm import embedder_factory
from rasa.shared.providers.embedding.embedding_client import EmbeddingClient
from rasa.shared.providers.embedding.embedding_response import EmbeddingResponse

# get the embedding client via factory
embedder: EmbeddingClient = embedder_factory(config, default_config)

# get the embedding response synchronously
sync_response: EmbeddingResponse = embedder.embed([doc_1, doc_2])
vectors: List[List[float]] = sync_response.data

# get the embedding response asynchronously
async_response: EmbeddingResponse = await embedder.aembed([doc_1, doc_2])
vectors: List[List[float]] = async_response.data

Changes to `invoke_llm`

The previous implementation of invoke_llm method in SingleStepLLMCommandGenerator, MultiStepLLMCommandGenerator, and the deprecated LLMCommandGenerator used llm_factory to instantiate LangChain clients. Since the factory now returns clients that conform to the new LLMClient protocol, any custom overrides of the invoke_llm method will need to be updated to accommodate the new return type.

Below you can find the invoke_llm method from Rasa Pro 3.9 and its updated version in Rasa Pro 3.10:

Rasa 3.9
Rasa 3.10

async def invoke_llm(self, prompt: Text) -> Optional[Text]:
    """Use LLM to generate a response.

    Args:
        prompt: The prompt to send to the LLM.

    Returns:
        The generated text.

    Raises:
        ProviderClientAPIException if an error during API call.
    """
    llm = llm_factory(self.config.get(LLM_CONFIG_KEY), DEFAULT_LLM_CONFIG)
    try:
        llm_response = await llm.acompletion(prompt)
        return llm_response.choices[0]
    except Exception as e:
        structlogger.error("llm_based_command_generator.llm.error", error=e)
        raise ProviderClientAPIException(
            message="LLM call exception", original_exception=e
        )

Changes to `SingleStepLLMCommandGenerator.predict_commands`

For SingleStepLLMCommandGenerator, the predict_commands method now includes a call to self._update_message_parse_data_for_fine_tuning(message, commands, flow_prompt). This function is essential for enabling the fine-tuning recipe.

If you have overridden the predict_commands method, you need to manually add this call to ensure proper functionality:

async def predict_commands(
        self,
        message: Message,
        flows: FlowsList,
        tracker: Optional[DialogueStateTracker] = None,
        **kwargs: Any,
    ) -> List[Command]:

    ...
    action_list = await self.invoke_llm(flow_prompt)
    commands = self.parse_commands(action_list, tracker, flows)

    self._update_message_parse_data_for_fine_tuning(message, commands, flow_prompt)

    return commands

Changes to the default configuration dictionary

The default configurations for the following components have been updated:

If you have custom implementations based on the default configurations for any of these components, ensure that your configuration dictionary aligns with the updates shown in the tables below, as the defaults have changed.

Default LLM configuration keys have been updated from:

DEFAULT_LLM_CONFIG = {
    "_type": "openai",
    "model_name": ...,
    "request_timeout": ...,
    "temperature": ...,
    "max_tokens": ...,
}

to:

DEFAULT_LLM_CONFIG = {
    "provider": "openai",
    "model": ...,
    "temperature": ...,
    "max_tokens": ...,
    "timeout": ...,
}

Similarly, default embedding configuration keys have been updated from:

DEFAULT_EMBEDDINGS_CONFIG = {
    "_type": "openai",
    "model": ...,
}

to:

DEFAULT_EMBEDDINGS_CONFIG = {
    "provider": "openai",
    "model": ...,
}

Be sure to update your custom configurations to reflect these changes in order to ensure continued functionality.

Dropped support for Python 3.8

Dropped support for Python 3.8 ahead of Python 3.8 End of Life in October 2024.

In Rasa Pro versions 3.10.0, 3.9.11 and 3.8.13, we needed to pin the TensorFlow library version to 2.13.0rc1 in order to remove critical vulnerabilities; this resulted in poor user experience when installing these versions of Rasa Pro with uv pip. Removing support for Python 3.8 will make it possible to upgrade to a stabler version of TensorFlow.

Rasa Pro 3.8 to Rasa Pro 3.9

LLMCommandGenerator

Starting from Rasa Pro 3.9 the former LLMCommandGenerator is replaced by SingleStepLLMCommandGenerator. The LLMCommandGenerator is now deprecated and will be removed in version 4.0.0.

The SingleStepLLMCommandGenerator differs from the LLMCommandGenerator in how it handles failures of the invoke_llm method. Specifically, if the invoke_llm method call fails in SingleStepLLMCommandGenerator, it raises a ProviderClientAPIException. In contrast, the LLMCommandGenerator simply returns None when the method call fails.

Slot Mappings

In case you had been using custom slot mapping type for slots set with the prediction of the LLM-based command generator, you need to update your assistant's slot configuration to use the new from_llm slot mapping type. Note that even if you have written custom slot validation actions (following the validate_<slot_name> convention) for slots set by the LLM-based command generator, you need to update your assistant's slot configuration to use the new from_llm slot mapping type.

For slots that are set only via a custom action e.g. slots set by external sources only, you must add the action name to the slot mapping:

slots:
  slot_name:
    type: text
    mappings:
      - type: custom
        action: custom_action_name

Rasa Pro 3.8.0 to Rasa Pro 3.8.1

Poetry Installation

Starting from Rasa Pro 3.8.1 in the 3.8.x minor series, we have upgraded the version Poetry for managing dependencies in the Rasa Pro Python package to 1.8.2. To install the latest micro versions of Rasa Pro in your project, you must first upgrade Poetry to version 1.8.2:

poetry self update 1.8.2

Rasa Pro 3.7 to 3.8

info

Starting from 3.8.0, Rasa and Rasa Plus have been merged into a single artifact, named Rasa Pro.

Installation

Following the merge we renamed the resulting python package and Docker image to rasa-pro.

Python package

Rasa Pro python package, for 3.8.0 and onward, is located at:

https://europe-west3-python.pkg.dev/rasa-releases/rasa-pro-python

Name of the package is rasa-pro.

Example of how to install the package:

pip install  --extra-index-url=https://europe-west3-python.pkg.dev/rasa-releases/rasa-pro-python/simple rasa-pro==3.8.0

While python package name was changed, the import process remains the same:

import rasa.core

from rasa import train

For more information on how to install Rasa Pro, please refer to the Python installation guide.

Helm Chart / Docker Image

Rasa Pro docker image, for 3.8.0 and onward, is located at:

europe-west3-docker.pkg.dev/rasa-releases/rasa-pro/rasa-pro

Example how to pull the image:

docker pull europe-west3-docker.pkg.dev/rasa-releases/rasa-pro/rasa-pro:3.8.0

For more information on how to install Rasa Pro Docker image, please refer to the Docker installation guide.

Component Yaml Configuration Changes

Follow the below instructions to update the configuration of Rasa Pro components in the 3.8 version:

ConcurrentRedisLockStore - update endpoints.yml to type: concurrent_redis:

lock_store:
  type: concurrent_redis

ContextualResponseRephraser- update endpoints.yml to either type: rephrase or type: rasa.core.ContextualResponseRephraser:

nlg:
  type: rephrase

Audiocodes and Vier CVG channels can be specified in credentials.yml using directly their channel name:

audiocodes:
  token: "sample_token"

vier_cvg:
  ...

EnterpriseSearchPolicy and IntentlessPolicy - update config.yml to only use the policy class name:

    policies:
      - name: EnterpriseSearchPolicy
      - name: IntentlessPolicy

Changes to default behaviour

info

With Rasa Pro 3.8, we introduced a couple of changes that rectifies the default behaviour of certain components. We believe these changes align better with the principles of CALM. If you are migrating an assistant built with Rasa Pro 3.7, please ensure you have checked if these changes affect your assistant.

Prompt Rendering

Rasa Pro 3.8 introduces a new feature flow-retrieval which ensures that only the flows that are relevant to the conversation context are included in the prompt sent to the LLM in the LLMCommandGenerator. This helps the assistant scale to a higher number of flows and also reduces the LLM costs.

This feature is enabled by default and we recommend to use it if the assistant has more than 40 flows. By default, the feature uses embedding models from OpenAI, but if you are using a different provider (for e.g. Azure), please ensure -

An embedding model is configured with the provider.
LLMCommandGenerator has been configured correctly to connect to the embedding provider. For example, see the section on configuration required to connect to Azure OpenAI service

If you wish to disable the feature you can configure the LLMCommandGenerator as:

config.yml

pipeline:
  - name: SingleStepLLMCommandGenerator
    ...
    flow_retrieval:
      active: false
    ...

Processing Chitchat

The default behaviour in Rasa Pro 3.7 to handle chitchat utterances was to rely on free form generative responses. This can lead to the assistant sending unwanted responses or responding to out of scope user utterances. The new default behaviour in Rasa Pro 3.8 is to rely on IntentlessPolicy to respond to chitchat utterances using pre-defined responses only.

If you were relying on free form generative responses to handle chitchat in Rasa Pro 3.7, you will now see a warning message when you train the same assistant with Rasa Pro 3.8 - " pattern_chitchat has an action step with action_trigger_chitchat, but IntentlessPolicy is not configured". This appears because the default definition of pattern_chitchat has been modified in Rasa Pro 3.8 to:

pattern_chitchat:
  description: handle interactions with the user that are not task-oriented
  name: pattern chitchat
  steps:
    - action: action_trigger_chitchat

For the assistant to be able to handle chitchat utterances, you have two options:

If you are happy with free-form generative responses for such user utterances, then you can override pattern_chitchat to:
pattern_chitchat:
description: handle interactions with the user that are not task-oriented
name: pattern chitchat
steps:
- action: utter_free_chitchat_response
If you want to switch to using pre-defined responses, you should first add IntentlessPolicy to the policies section of the config -
```yaml
policies:
- name: IntentlessPolicy
```

Next, you should add response templates for the pre-defined responses you want the assistant to consider when responding to a chitchat user utterance.

Handling of categorical slots

Rasa Pro versions <= 3.7.8 used to store the value of a categorical slot in the same casing as it was either specified in the user message or predicted by the LLM in a SetSlot command. This wasn't necessarily same as the casing used in the corresponding possible value defined for that slot in the domain. For e.g, if the categorical slot was defined to have [A, B, C] as the possible values and the prediction was to set it to a then the slot would be set to a. This lead to problems downstream when that slot had to be used in other primitives i.e. flows or custom action.

Rasa Pro 3.7.9 fixes this by always storing the slot value in the same casing as defined in the domain. So, in the above example, the slot would now be stored as A instead of a. This ensures that the user is writing business logic for slot comparisons, for e.g. if conditions in flows, using the same casing as defined by them in the domain.

If you are migrating from Rasa pro versions <= 3.7.8, please double check your flows and custom actions to make sure none of them break because of this change.

Update default signature of LLM calls

In Rasa Pro >= 3.8 we switched from doing synchronous LLM calls to asynchronous calls. We updated all components that use an LLM, e.g.

LLMCommandGenerator
ContextualResponseRephraser
EnterpriseSearchPolicy
IntentlessPolicy

This can potentially break assistants migrating to 3.8 that have sub-classed one of these components in their own custom components.

For example, the method predict_commands in the LLMCommandGenerator is now async and needs to await the methods _generate_action_list_using_llm and flow_retrieval.filter_flows as these methods are also async. For more information on asyncio please check their documentation.

Dependency Upgrades

We've updated our core dependencies to enhance functionality and performance across our platform.

Spacy 3.7.x

Upgraded from >=3.6 to >=3.7.

We have transitioned to using Spacy version 3.7.x to benefit from the latest enhancements in natural language processing. If you're using any spacy models with your assistant, please update them to Spacy 3.7.x compatible models.

Pydantic 2.x

Upgraded from >=1.10.9,<1.10.10 to ^2.0.

Along with the Spacy upgrade, we have moved to Pydantic version 2.x, which necessitates updates to Pydantic models. For assistance with updating your models, please refer to the Pydantic Migration Guide. This ensures compatibility with the latest improvements in data validation and settings management.

Rasa Pro 3.7.9 to Rasa Pro 3.7.10

Poetry Installation

Starting from Rasa Pro 3.7.10 in the 3.7.x minor series, we have upgraded the version Poetry for managing dependencies in the Rasa Pro Python package to 1.8.2. To install Rasa Pro in your project, you must first upgrade Poetry to version 1.8.2:

poetry self update 1.8.2

Rasa Pro 3.7.8 to Rasa Pro 3.7.9

Changes to default behaviour

Handling of categorical slots

If you are migrating from Rasa pro versions <= 3.7.8, please double check your flows and custom actions to make sure none of them break because of this change.

Rasa 3.6 to Rasa Pro 3.7

Installation

info

Starting from Rasa 3.7.0, Rasa has moved to a new package registry and Docker registry. You will need to update your package registry to install Rasa 3.7.0 and later versions. If you are a Rasa customer, please reach out to your Rasa account manager or support obtain a license.

Python package

Rasa python package for 3.7.0 has been moved to python package registry.

https://europe-west3-python.pkg.dev/rasa-releases/rasa-plus-py

Name of the package is rasa.

Example of how to install the package:

pip install  --extra-index-url=https://europe-west3-python.pkg.dev/rasa-releases/rasa-plus-py/simple rasa==3.7.0

For more information on how to install Rasa Pro, please refer to the Python installation guide.

Helm Chart / Docker Image

Rasa docker image for 3.7.0 is located at:

europe-west3-docker.pkg.dev/rasa-releases/rasa-docker/rasa

Example how to pull the image:

docker pull europe-west3-docker.pkg.dev/rasa-releases/rasa-docker/rasa:3.7.0

For more information on how to install Rasa Pro Docker image, please refer to the Docker installation guide.

Rasa 3.0 to 3.1

Machine Learning Components

TensorFlow Upgrade

Due to the TensorFlow upgrade, we can't guarantee the exact same output and hence model performance if your configuration uses LanguageModelFeaturizer. This applies to the case where the model is re-trained with the new Rasa version without changing the configuration, random seeds, and data as well as to the case where a model trained with a previous version of Rasa is loaded with this new version for inference.

Please check whether your trained model still performs as expected and retrain if needed.

NLU JSON Format

NLU training data in JSON format is deprecated and will be removed in Rasa 4.0. Please use rasa data convert nlu -f yaml --data <path to NLU data> to convert your NLU JSON data to YAML format before support for NLU JSON data is removed.

Rasa 2.x to 3.0

Markdown Data

Markdown is no longer supported — all the supporting code that was previously deprecated is now removed, and the convertors are removed as well.

The related CLI commands rasa data convert responses and rasa data convert config were removed.

If you still have training data in Markdown format then the recommended approach is to use Rasa 2.x to convert your data from Markdown to YAML. Please use the commands described here.

Model Configuration

It is required to specify the used recipe within the model configuration. As of now Rasa only supports the default.v1 recipe and will continue using it even if you don't specify a recipe in the model configuration. To avoid breaking changes in the future you should to specify recipe: "default.v1" at the top of your model configuration:

Rasa 2.0 (old)
Rasa 3.0 (new)

recipe: default.v1

language: en

pipeline:
  ...
policies:
  ...

Custom Policies and Custom Components

Rasa 3.0 changed the way NLU components and policies are trained and run during inference. As part of these changes the interfaces for NLU components and policies have been unified and adapted.

The next sections outline which adaptions are required to run your custom NLU components and policies with Rasa 3.0.

caution

Please read the updated guide on custom graph components here before continuing to follow the step-by-step guide to migrate your own custom graph components.

Type Annotations

Until Rasa 3.0 type annotations were not required in custom policies or custom NLU components. It is now required to use type annotations in custom NLU components and policies. Rasa uses these type annotations to validate that your graph components are compatible and correctly configured. As outlined in the custom components guide it is not allowed to use forward references.

Forward References with Python 3.7

Use the from __future__ import annotations import to avoid using forward references with Python 3.7.

Rasa 2.0 (old)
Rasa 3.0 (new)

from __future__ import annotations
from typing import List, Any

from rasa.core.policies.policy import Policy
from rasa.engine.storage.resource import Resource
from rasa.shared.core.domain import Domain
from rasa.shared.core.generator import TrackerWithCachedStates

class MyPolicy(Policy):
    def train(
        self,
        training_trackers: List[TrackerWithCachedStates],
        domain: Domain,
        **kwargs: Any,
    ) -> Resource:
        ...

Changes to Custom NLU Components

Inheriting from GraphComponent

NLU components which previously inherited from one of the following classes additionally need to inherit from the GraphComponent interface:

SparseFeaturizer
DenseFeaturizer
IntentClassifier
EntityExtractor
Component

This snippet shows the required changes:

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.engine.graph import GraphComponent
from rasa.nlu.featurizers.sparse_featurizer.sparse_featurizer import SparseFeaturizer

class MyNLUComponent(GraphComponent, SparseFeaturizer):
    ...

Inheriting from EntityExtractorMixin instead of EntityExtractor

The EntityExtractor class was renamed to EntityExtractorMixin:

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.engine.graph import GraphComponent
from rasa.nlu.extractors.extractor import EntityExtractorMixin

class MyNLUComponent(GraphComponent, EntityExtractorMixin):
    ...

Instantiating a NLU Component for Training

NLU components are no longer instantiated via their constructor. Instead, all NLU components have to override the create method of the GraphComponent interface. The passed in configuration is your NLU component's default configuration including any updates from your model configuration file.

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import Dict, Text, Any

from rasa.engine.graph import GraphComponent, ExecutionContext
from rasa.engine.storage.resource import Resource
from rasa.engine.storage.storage import ModelStorage
from rasa.nlu.classifiers.classifier import IntentClassifier

class MyNLUComponent(GraphComponent, IntentClassifier):

  def __init__(self, component_config: Dict[Text, Any]) -> None:
      self.component_config = component_config
    ...

  @classmethod
  def create(
      cls,
      config: Dict[Text, Any],
      model_storage: ModelStorage,
      resource: Resource,
      execution_context: ExecutionContext,
  ) -> GraphComponent:
      return cls(config)

Persisting a Trained NLU Component

NLU components used to be persisted by a call to the NLU component's persist method from outside the NLU component itself. With Rasa 3.0 NLU components are responsible for persisting themselves. Use the provided model_storage and resource parameters to persist your NLU component at the end of the training and then return the resource as result of your NLU component's train method. See component persistence for more details.

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.engine.graph import GraphComponent
from rasa.engine.storage.resource import Resource
from rasa.nlu.classifiers.classifier import IntentClassifier
from rasa.shared.nlu.training_data.training_data import TrainingData

class MyNLUComponent(GraphComponent, IntentClassifier):
    def train(self, training_data: TrainingData) -> Resource:
        ...
        self.persist()
        return self._resource

    def persist(self) -> None:
        with self._model_storage.write_to(self._resource) as directory:
            model_data_file = directory / "model_data.json"
            rasa.shared.utils.io.dump_obj_as_json_to_file(model_data_file,
                self.get_model_data())
    ...

Instantiating a Trained NLU Component

Previously NLU components had to persist their own configuration. Now the config passed into load will automatically contain the configuration which your model was trained with. To instantiate a persisted NLU component, you need to use model_storage and resource in your NLU component's load method.

Rasa 2.0 (old)
Rasa 3.0 (new)

from __future__ import annotations
import json
from typing import Any, Text, Dict

from rasa.engine.graph import GraphComponent, ExecutionContext
from rasa.engine.storage.resource import Resource
from rasa.engine.storage.storage import ModelStorage
from rasa.nlu.classifiers.classifier import IntentClassifier
from rasa.shared.exceptions import FileIOException

class MyNLUComponent(GraphComponent, IntentClassifier):
    @classmethod
    def load(
        cls,
        config: Dict[Text, Any],
        model_storage: ModelStorage,
        resource: Resource,
        execution_context: ExecutionContext,
        **kwargs: Any,
    ) -> MyNLUComponent:
        model_data = {}

        try:
            with model_storage.read_from(resource) as path:

                model_data_file = path / "model_data.json"
                model_data = json.loads(rasa.shared.utils.io.read_file(model_data_file))

        except (ValueError, FileNotFoundError, FileIOException):
            logger.debug(
                f"Couldn't load metadata for component '{cls.__name__}' as the persisted "
                f"model data couldn't be loaded."
            )

        return cls(
            config, model_data=model_data
        )

Providing a Default Configuration for an NLU Component

The default configuration is no longer a static class property but instead returned by the static method get_default_config:

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import Text, Any, Dict

from rasa.engine.graph import GraphComponent
from rasa.nlu.classifiers.classifier import IntentClassifier

class MyNLUComponent(GraphComponent, IntentClassifier):
    ...
    @staticmethod
    def get_default_config() -> Dict[Text, Any]:
        return {"key1": "value1"}

Augmenting Training Data in an NLU Component

NLU Components like tokenizers or featurizers augment the training data with their output during the model training. Their output is required by NLU components later in the pipeline. Typically, featurizers require tokenized messages and intent classifiers require featurized training data to train themselves. Rasa 3.0 makes these different purposes explicit. Previously both NLU component training and training data augmentation were done as part of the train method. In Rasa 3.0 they are split into train and process_training_data:

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.engine.graph import GraphComponent
from rasa.engine.storage.resource import Resource
from rasa.nlu.featurizers.sparse_featurizer.sparse_featurizer import SparseFeaturizer
from rasa.shared.nlu.training_data.training_data import TrainingData

class MyNLUComponent(GraphComponent, SparseFeaturizer):
    def train(self, training_data: TrainingData) -> Resource:
        self.train_featurizer(training_data)

        self.persist()
        return self._resource

    def process_training_data(self, training_data: TrainingData) -> TrainingData:
        for message in training_data.training_examples:
            self.add_features(message)

        return training_data

Handling Lists of Messages During Inference in an NLU Component

NLU components used to receive a single Message object during inference. Starting with Rasa 3.0 all NLU components have to support a list of messages during inference. Unless your component supports batch predictions the easiest way to handle this is to loop over the messages. It is also required to return the message objects at the end of the process method.

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import List

from rasa.engine.graph import GraphComponent
from rasa.nlu.classifiers.classifier import IntentClassifier
from rasa.shared.nlu.training_data.message import Message

class MyNLUComponent(GraphComponent, IntentClassifier):
    def process(self, messages: List[Message]) -> List[Message]:
        for message in messages:
            self.predict(message)

        return messages

Registering your NLU Component

Before you can use your custom NLU component you have to register your NLU component using the DefaultV1Recipe.register decorator. The NLU component types correspond to the existing parent classes:

Tokenizer: ComponentType.MESSAGE_TOKENIZER
SparseFeaturizer / DenseFeaturizer: ComponentType.MESSAGE_FEATURIZER
IntentClassifier: ComponentType.INTENT_CLASSIFIER
EntityExtractor: ComponentType.ENTITY_EXTRACTOR
If your NLU component provides a pretrained model which should be used by other NLU components during training and inference use ComponentType.MODEL_LOADER

Specify is_trainable=True if the train method of your component should be called during training.

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.engine.recipes.default_recipe import DefaultV1Recipe
from rasa.nlu.classifiers.classifier import IntentClassifier
from rasa.engine.graph import GraphComponent

@DefaultV1Recipe.register(
    DefaultV1Recipe.ComponentType.INTENT_CLASSIFIER, is_trainable=True
)
class MyNLUComponent(GraphComponent, IntentClassifier):
    ...

Using a Model Provider with your NLU Component

If your NLU component requires a pretrained model such as a Spacy or Mitie language model you have to specify the NLU component which provides this model in your model's pipeline before the NLU component which requires the model. In addition to this you now also need to specify the model loading component in the model_from parameter in the register decorator. The model will then be passed to your model's train, process_training_data and process methods:

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import List

from rasa.engine.graph import GraphComponent
from rasa.engine.recipes.default_recipe import DefaultV1Recipe
from rasa.engine.storage.resource import Resource
from rasa.nlu.classifiers.classifier import IntentClassifier
from rasa.nlu.utils.spacy_utils import SpacyModel
from rasa.shared.nlu.training_data.message import Message
from rasa.shared.nlu.training_data.training_data import TrainingData

@DefaultV1Recipe.register(
    DefaultV1Recipe.ComponentType.INTENT_CLASSIFIER, is_trainable=True, model_from="SpacyNLP"
)
class MyNLUComponent(GraphComponent, IntentClassifier):
    def train(
        self, training_data: TrainingData, model: SpacyModel) -> Resource:
        spacy_nlp = model.model
        ...

    def process(self, messages: List[Message], model: SpacyModel) -> List[Message]:
        spacy_nlp = model.model
        ...

Changes to Custom Policies

This guide leads you through the migration of a custom policy step by step.

Instantiating a Policy for Training

Policies are no longer instantiated via their constructor. Instead, all policies have to implement a create method. During the policy instantiation the configuration from the model configuration is passed in as a dictionary instead of as separate parameters. Similarly, thefeaturizers are no longer instantiated outside of policies. Instead, the super class rasa.core.policies.policy.Policy instantiates the featurizers itself.

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import Optional, Dict, Text, Any

from rasa.core.featurizers.tracker_featurizers import TrackerFeaturizer
from rasa.core.policies.policy import Policy
from rasa.engine.graph import ExecutionContext
from rasa.engine.storage.resource import Resource
from rasa.engine.storage.storage import ModelStorage

class MyPolicy(Policy):
    def __init__(
        self,
        config: Dict[Text, Any],
        model_storage: ModelStorage,
        resource: Resource,
        execution_context: ExecutionContext,
        featurizer: Optional[TrackerFeaturizer] = None,
    ) -> None:
        super().__init__(
          config, model_storage, resource, execution_context, featurizer
        )
        ...
    ...

  @classmethod
  def create(
      cls,
      config: Dict[Text, Any],
      model_storage: ModelStorage,
      resource: Resource,
      execution_context: ExecutionContext,
  ) -> MyPolicy:
      return cls(config, model_storage, resource, execution_context)

Persisting a Trained Policy

Policies used to be persisted by a call to the policy's persist method from outside the policy itself. With Rasa 3.0 policies are responsible for persisting themselves. Use the provided model_storage and resource parameters to persist your graph component at the end of the training and then return the resource as result of your policy's train method. See graph component persistence for more details.

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import List, Any

from rasa.core.policies.policy import Policy
from rasa.engine.storage.resource import Resource
from rasa.shared.core.domain import Domain
from rasa.shared.core.generator import TrackerWithCachedStates

class MyPolicy(Policy):
    def train(
        self,
        training_trackers: List[TrackerWithCachedStates],
        domain: Domain,
        **kwargs: Any,
    ) -> Resource:
        ...
        self.persist()
        return self._resource

    def persist(self) -> None:
        with self._model_storage.write_to(self._resource) as directory:
            if self.featurizer is not None:
                self.featurizer.persist(directory)

            file_path = directory / "model_data.json"
            rasa.shared.utils.io.dump_obj_as_json_to_file(file_path,
                self.get_model_data())
    ...

Instantiating a Trained Policy

Previously policies had to persist their own configuration. Now the config passed into load will automatically contain the configuration which your model was trained with.

To instantiate a persisted policy, you need to use model_storage and resource in your policy's load method.

Rasa 2.0 (old)
Rasa 3.0 (new)

import json
from typing import Dict, Text, Any

from rasa.core.featurizers.tracker_featurizers import TrackerFeaturizer
from rasa.core.policies.policy import Policy
from rasa.engine.graph import ExecutionContext
from rasa.engine.storage.resource import Resource
from rasa.engine.storage.storage import ModelStorage
from rasa.shared.exceptions import FileIOException

class MyPolicy(Policy):
    @classmethod
    def load(
        cls,
        config: Dict[Text, Any],
        model_storage: ModelStorage,
        resource: Resource,
        execution_context: ExecutionContext,
        **kwargs: Any,
    ) -> MyPolicy:
        featurizer = None
        model_data = {}

        try:
            with model_storage.read_from(resource) as path:
                if (Path(path) / FEATURIZER_FILE).is_file():
                    featurizer = TrackerFeaturizer.load(path)

                model_data_file = path / "model_data.json"
                model_data = json.loads(rasa.shared.utils.io.read_file(model_data_file))

        except (ValueError, FileNotFoundError, FileIOException):
            logger.debug(
                f"Couldn't load metadata for policy '{cls.__name__}' as the persisted "
                f"metadata couldn't be loaded."
            )

        return cls(
            config, model_storage, resource, execution_context,
            featurizer=featurizer, model_data=model_data
        )

Providing a Default Configuration for a Policy

The default configuration is no longer provided via default values in your policy's constructor but instead returned by the static method get_default_config:

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import Dict, Text, Any
from rasa.core.policies.policy import Policy

class MyPolicy(Policy):

    def __init__(self, config: Dict[Text, Any]) -> None:
        ...

    @staticmethod
    def get_default_config() -> Dict[Text, Any]:
        return {"key1": "value1"}

Using End-To-End Features in a Policy

To use a custom end-to-end policy in Rasa Open Source 2, you had to use the interpreter parameter to featurize the tracker events manually. In Rasa 3.0, you need to register a policy that requires end-to-end features with type ComponentType.POLICY_WITH_END_TO_END_SUPPORT. The features will be precomputed and passed into your policy during training and inference.

caution

End-To-End features will only be computed and provided to your policy if your training data actually contains end-to-end training data.

Rasa 2.0 (old)
Rasa 3.0 (new)

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

from rasa.core.featurizers.precomputation import MessageContainerForCoreFeaturization
from rasa.core.policies.policy import PolicyPrediction, Policy
from rasa.engine.recipes.default_recipe import DefaultV1Recipe
from rasa.engine.storage.resource import Resource
from rasa.shared.core.domain import Domain
from rasa.shared.core.generator import TrackerWithCachedStates
from rasa.shared.core.trackers import DialogueStateTracker

@DefaultV1Recipe.register(
    DefaultV1Recipe.ComponentType.POLICY_WITH_END_TO_END_SUPPORT, is_trainable=True
)
class MyPolicy(Policy):
    def train(
        self,
        training_trackers: List[TrackerWithCachedStates],
        domain: Domain,
        precomputations: Optional[MessageContainerForCoreFeaturization] = None,
    ) -> Resource:
        ...
        model_data, label_ids = self._prepare_for_training(
          training_trackers, domain, precomputations,
        )
        ...

    def predict_action_probabilities(
        self,
        tracker: DialogueStateTracker,
        domain: Domain,
        precomputations: Optional[MessageContainerForCoreFeaturization] = None,
        rule_only_data: Optional[Dict[Text, Any]] = None,
        **kwargs: Any,
    ) -> PolicyPrediction:
        ...
        tracker_state_features = self._featurize_tracker(
            tracker, domain, precomputations, rule_only_data=rule_only_data
        )
        ...

Registering a Policy

Before you can use your custom policy you have to register your policy using the DefaultV1Recipe.register decorator. If your policy requires end-to-end features specify the graph component type POLICY_WITH_END_TO_END_SUPPORT. Otherwise, use POLICY_WITHOUT_END_TO_END_SUPPORT. Specify is_trainable=True if the train method of your policy should be called during the training. If your policy is only used during inference use is_trainable=False.

Rasa 2.0 (old)
Rasa 3.0 (new)

from rasa.core.policies.policy import Policy
from rasa.engine.recipes.default_recipe import DefaultV1Recipe

@DefaultV1Recipe.register(
    DefaultV1Recipe.ComponentType.POLICY_WITH_END_TO_END_SUPPORT,
    is_trainable=True
)
class MyPolicy(Policy):
    ...

Providing Rule-only Data to a Policy

Rasa allows excluding forms or slots which are completely handled by rules from becoming features in other policies. In Rasa 2 this information was passed onto the policies using the set_shared_policy_states method which set the policy attribute _rule_only_data. Rasa passes the names of rule-only slots and forms via the predict_action_probabilities method. The passed rule_only_data can be None in case the RulePolicy is not part of your model configuration.

Rasa 2.0 (old)
Rasa 3.0 (new)

from typing import Optional, Dict, Text, Any

from rasa.core.policies.policy import Policy, PolicyPrediction
from rasa.shared.core.domain import Domain
from rasa.shared.core.trackers import DialogueStateTracker

class MyPolicy(Policy):
    def predict_action_probabilities(
        self,
        tracker: DialogueStateTracker,
        domain: Domain,
        rule_only_data: Optional[Dict[Text, Any]] = None,
  ) -> PolicyPrediction:
      ...

Training data

Upgrading `version` from `2.0` to `3.0`

At the top of your training data files, you need to change version: "2.0" to version: "3.1".

We follow semantic versioning for training data versions. This means breaking changes result in a new major version, while new features result in a new minor version. The latest training data version is 3.1.

The improvements to slot mappings in Rasa 3.0 were breaking changes, so we needed to upgrade from major version 2.0 to major version 3.0.

`TrainingDataImporter`

TrainingDataImporter and all its implementations are updated to contain only synchronous methods. If you have a custom data importer or rely on some functions provided by TrainingDataImporter, you need to update your implementation and function calls.

For example, this is how data loading should look like in Rasa 3.0:

from typing import Text
from rasa.shared.importers.importer import TrainingDataImporter

def load_data(domain_path: Text, config_path: Text):
    file_importer = TrainingDataImporter.load_from_config(
        config_path, domain_path
    )
    # note that all the functions below were async before:
    config = file_importer.get_config()
    domain = file_importer.get_domain()
    stories = file_importer.get_stories()
    nlu_data = file_importer.get_nlu_data()

Since any custom importer implements TrainingDataImporter, you should update your custom importer to contain only sync methods as well:

from typing import Dict

from rasa.shared.core.domain import Domain
from rasa.shared.importers.importer import TrainingDataImporter


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

    # this function was async before
    def get_domain(self) -> Domain:
        pass

    # this function was also async before
    def get_config(self) -> Dict:
        pass

    # ...

template_variables and e2e arguments also got removed from get_stories method of TrainingDataImporter. Its new signature looks this way:

from typing import Optional

from rasa.shared.nlu.interpreter import RegexInterpreter
from rasa.shared.core.training_data.structures import StoryGraph

class TrainingDataImporter:
    # ...

    def get_stories(
        self,
        interpreter: "NaturalLanguageInterpreter" = RegexInterpreter(),
        exclusion_percentage: Optional[int] = None,
    ) -> StoryGraph:
        pass

    # ...

Training

`rasa train --dry-run`

Due to changes in the model architecture the behavior of rasa train --dry-run changed. The exit codes now have the following meaning:

0 means that the model does not require an expensive retraining. However, the responses might still require updating by running rasa train
1 means that one or multiple components require to be retrained.
8 means that the --force flag was used and hence any cached results are ignored and the entire model is retrained.

Machine Learning Components

Normalization of Confidences in `DIETClassifier` and `ResponseSelector`

DIETClassifier and ResponseSelector will no longer automatically report re-normalized confidences when ranking_length is set to a value greater than 0. This change affects the reported confidences but does not influence the final predicted intent, which might be used by policies. However, since the reported confidences are affected you might have to tune the thresholds for fallback mechanisms again. The previous behavior can still be enforced by setting renormalize_confidences=True when using model_confidence=softmax.

Normalization of confidences in `TEDPolicy`

Predictions of TEDPolicy will no longer be modified by masking and renormalizing confidences. This change can affect the maximum confidence predicted by the TEDPolicy and thereby affect the final result of the policy ensemble. However, the previous behavior can still be enforced by setting ranking_length=10 and renormalize_confidences=True.

Removed Policies

Several dialogue policies that were deprecated in Rasa 2.x have been removed in Rasa 3.0. If you are migrating a config file with a removed policy, consult the following migration guides for the individual policies:

FallbackPolicy migration guide
TwoStageFallbackPolicy migration guide
MappingPolicy migration guide
FormPolicy migration guide
SklearnPolicy should be replaced with TEDPolicy. It is recommended to use the default TEDPolicy config as a starting point.

Removed Tokenizers and Featurizers

The ConveRTTokenizer, LanguageModelTokenizer, and HFTransformersNLP featurizer components were deprecated in Rasa 2.x and have been removed in Rasa 3.0. See the migration guide for Rasa 2.x for replacing these components in your pipeline.

Slot Mappings

As of version 3.0, there is a single explicit mechanism for slot filling enabled by defining slot mappings for each slot in the slots section of the domain file. This approach keeps slots up to date over the course of a conversation, and removes duplicated effort in mapping the same slots in multiple forms. It is still possible to fill slots from arbitrary custom actions and not update them on every turn of the conversation if that behavior is desired.

This new mechanism replaces the implicit slot setting via auto-fill of slots with entities of the same name. The auto_fill key in the domain is no longer available, as well as the auto_fill parameter in the constructor of the Slot class.

While forms continue to request the next slot, slot extraction is now delegated to the default action action_extract_slots. This action runs in the background automatically after each user turn. Like action_listen, it should not be included in stories.

Each slot in the slots section must include the mappings key. The same keys used for predefined mappings in 2.0 are available in 3.0. Additionally, you can define slots with custom mappings implemented in a custom action which will be run on every user turn, for example:

slots:
  is_existing_customer:
    type: bool
    mappings:
      - type: custom
        action: action_verify_customer_status

You can use slot validation actions to either validate slots with predefined mappings, or to both extract and validate slots with custom mappings.

Slots which will be filled by arbitrary custom actions in the course of the conversation, and which should not be updated on every user turn, should be listed with mappings of type custom and no action. For example:

slots:
  handoff_completed:
    type: bool
    initial_value: false
    mappings:
      - type: custom

This slot's value will only change when a custom action is predicted that sets it. This mapping maintains the behavior from 2.x for a slot which was not filled by an entity or by slot mappings in a form.

note

The required_slots of a form used to be a list of slot mappings. Since slot mappings are relocated to the slots section of the domain, required_slots has been converted to a list of slot names only.

Automatic migration from 2.0 domain format to the 3.0 format

The only data file that has changed in format is the domain file. To migrate automatically to the 3.0 domain format, you can run the following command:

rasa data migrate -d DOMAIN --out OUT_PATH

In addition to creating a valid 3.0 domain in the indicated out path, this command will automatically backup your original domain file(s) in a file labeled original_domain.yml or original_domain directory if a directory was provided instead.

To maintain the behavior of forms in the 2.0 format, all migrated slot mappings will include mapping conditions for each form. This can be changed manually according to your use case. See the docs on mapping conditions for more information.

Manually migrating from 2.0 domain format to the 3.0 format

Each slot in the slots section of the domain will need a new key mappings. This key is a list of mappings moved from forms, while the required_slots field collapses to a list of slot names.

Let's consider the following 2.0 domain file:

entities:
  - cuisine
  - number
slots:
  cuisine:
    type: text
  num_people:
    type: float
  outdoor_seating:
    type: bool
forms:
  restaurant_form:
    required_slots:
      cuisine:
        - type: from_entity
          entity: cuisine
      num_people:
        - type: from_entity
          entity: number
      outdoor_seating:
        - type: from_intent
          intent: affirm
          value: true
        - type: from_intent
          intent: deny
          value: false

The initial result of migrating this domain to 3.0 format would look like this:

entities:
  - cuisine
  - number
slots:
  cuisine:
    type: text
    mappings:
      - type: from_entity
        entity: cuisine
  num_people:
    type: float
    mappings:
      - type: from_entity
        entity: number
  outdoor_seating:
    type: bool
    mappings:
      - type: from_intent
        intent: affirm
        value: true
      - type: from_intent
        intent: deny
        value: false
forms:
  restaurant_form:
    required_slots:
      - cuisine
      - num_people
      - outdoor_seating

For slots that should be filled only in the context of a form, add mapping conditions to specify which form(s) should be active, as well as indicate if the requested_slot should be the same slot. Adding conditions is required to preserve the behavior of slot mappings from 2.0, since without them the mappings will be applied on each user turn regardless of whether a form is active or not.

slots:
  outdoor_seating:
    type: bool
    mappings:
      - type: from_intent
        intent: affirm
        value: true
        conditions:
          - active_loop: restaurant_form
            requested_slot: outdoor_seating
      - type: from_intent
        intent: deny
        value: false
        conditions:
          - active_loop: restaurant_form
            requested_slot: outdoor_seating

Rasa-SDK Modifications

If you have used FormValidationAction to define custom extraction and validation code in which you override the required_slots method, note that slots_mapped_in_domain argument has been replaced by the domain_slots argument. You must make this replacement to continue using your custom code.

If you have been dynamically filling slots not present in the form's required_slots defined in the domain.yml file, note that this behaviour is no longer supported in 3.x. Any dynamic slots with custom mappings, which are set in the last user turn, will be filled only if they are returned by the required_slots method of the custom action inheriting from FormValidationAction. To maintain the 2.x behaviour, you must now override the required_slots method of this custom action as per the strong recommendation listed in the dynamic form documentation.

To extract custom slots that are not defined in any form's required_slots, you should now use a global custom slot mapping and extend the ValidationAction class.

note

If you have custom validation actions extending FormValidationAction which override required_slots method, you should double-check the dynamic form behavior of your migrated assistant. Slots set by the default action action_extract_slots may need to be reset within the context of your form by the custom validation actions for the form's required slots. For example, if your form dynamically adds a required slot after the first slot is filled, you may want to reset the potential required slot as part of the first required slot's validation method to ensure it will be empty when added.

Rasa 2.7 to 2.8

caution

This release breaks backward compatibility of machine learning models. It is not possible to load models trained with previous versions of Rasa. Please re-train your assistant before using this version.

Deprecations

Tracker Featurizers

training_states_actions_and_entities method of TrackerFeaturizer, FullDialogueTrackerFeaturizer and MaxHistoryTrackerFeaturizer classes is deprecated and will be removed in Rasa 3.0 . If you had a custom tracker featurizer which relied on this method from any of the above classes, please use training_states_labels_and_entities instead.

training_states_and_actions method of TrackerFeaturizer, FullDialogueTrackerFeaturizer and MaxHistoryTrackerFeaturizer classes is deprecated and will be removed in Rasa 3.0 . If you had a custom tracker featurizer which relied on this method from any of the above classes, please use training_states_and_labels instead.

State Featurizer

encode_all_actions method of SingleStateFeaturizer class is deprecated and will be removed in Rasa 3.0 . It is recommended to use the method encode_all_labels instead.

Incremental Training

Users don't need to specify an additional buffer size for sparse featurizers anymore during incremental training.

Space for new sparse features are created dynamically inside the downstream machine learning models - DIETClassifier, ResponseSelector. In other words, no extra buffer is created in advance for additional vocabulary items and space will be dynamically allocated for them inside the model.

This means there's no need to specify additional_vocabulary_size for CountVectorsFeaturizer or number_additional_patterns for RegexFeaturizer. These parameters are now deprecated.

Before

pipeline:
  - name: "WhitespaceTokenizer"
  - name: "RegexFeaturizer"
    number_additional_patterns: 100
  - name: "CountVectorsFeaturizer"
    additional_vocabulary_size: { text: 100, response: 20 }

Now

pipeline:
  - name: "WhitespaceTokenizer"
  - name: "RegexFeaturizer"
  - name: "CountVectorsFeaturizer"

Machine Learning Components

The option model_confidence=linear_norm is deprecated and will be removed in Rasa 3.0.0.

Rasa 2.3.0 introduced linear_norm as a possible value for model_confidence parameter in machine learning components such as DIETClassifier, ResponseSelector and TEDPolicy. Based on user feedback, we have identified multiple problems with this option. Therefore, model_confidence=linear_norm is now deprecated and will be removed in Rasa 3.0.0. If you were using model_confidence=linear_norm for any of the mentioned components, we recommend to revert it back to model_confidence=softmax and re-train the assistant. After re-training, we also recommend to re-tune the thresholds for fallback components.

Rasa 2.5 to 2.6

Forms

New `ignored_intents` parameter in Forms

There is a new parameter under Forms called ignored_intents. This parameter can be used to prevent any required slots in a form from being filled with the specified intent or intents. Please see the Forms documentation for examples and more information on how to use it in your domain.yml file.

Before, if a user did not want to fill any slots of a form with a specified intent they would have to define it under the not_intent parameter for every slot mapping as shown in the following example :

domain.yml

forms:
  restaurant_form:
      cuisine:
      - type: from_entity
        entity: cuisine
        not_intent: chitchat
      num_people:
      - type: from_entity
        entity: number
        intent: [inform, request_restaurant]
        not_intent: chitchat
      feedback:
      - type: from_entity
        entity: feedback
        not_intent: chitchat

By introducing the ignored_intents parameter, we now only need to define it in one place and it will affect all the slots of the form :

domain.yml

forms:
  restaurant_form:
    ignored_intents: chitchat
    required_slots:
      cuisine:
      - type: from_entity
        entity: cuisine
      num_people:
      - type: from_entity
        entity: number
        intent: [inform, request_restaurant]
      feedback:
      - type: from_entity
        entity: feedback
      - type: from_text

Rasa 2.4 to 2.5

Machine Learning Components

`DIET`, `TED`, and `ResponseSelector`

The former weight_sparsity parameter of the DIETClassifier, TEDPolicy, and the ResponseSelector, is now deprecated and superseded by the new connection_density parameter. The old weight_sparsity is roughly equivalent to 1 - connection_density, except at very low densities (high sparsities).

To avoid deprecation issues, you should set connection_density to 1 - your former weight_sparsity setting throughout the config file. (If you left weight_sparsity at its default setting, you don't need to do anything.)

SpaCy 3.0

Rasa now supports spaCy 3.0. This means that we can support more features for more languages but this also introduced a breaking change. SpaCy 3.0 deprecated the spacy link <language model> command. So from now on you need to use the the full model name in the config.yml file.

Before

Before you could run spacy link en en_core_web_md and then we would be able to pick up the correct model from the language parameter.

language: en

pipeline:
  - name: SpacyNLP

Now

This behavior will be deprecated and instead you'll want to be explicit in config.yml.

language: en

pipeline:
  - name: SpacyNLP
    model: en_core_web_md

Fallback

To make the transition easier, Rasa will try to fall back to a medium spaCy model whenever a compatible language is configured for the entire pipeline in config.yml, even if you don't specify a model. This fallback behavior is temporary and will be deprecated in Rasa 3.0.0.

We've updated our docs to reflect these changes. All examples now show a direct link to the correct spaCy model. We've also added a warning to the SpaCyNLP docs that explains the fallback behavior.

Rasa 2.3 to Rasa 2.4

Deprecating `template` for `response`

NLG Server

Changed request format to send response as well as template as a field. The template field will be removed in Rasa 3.0.0.

rasa.core.agent

The terminology template is deprecated and replaced by response. Support for template from the NLG response will be removed in Rasa 3.0.0. Please see here for more details.

rasa.core.nlg.generator

generate() now takes in utter_action as a parameter.
The terminology template is deprecated and replaced by response. Support for template in the NaturalLanguageGenerator will be removed in Rasa 3.0.0.

rasa.shared.core.domain

The property templates is deprecated. Use responses instead. It will be removed in Rasa 3.0.0.
retrieval_intent_templates will be removed in Rasa 3.0.0. Please use retrieval_intent_responses instead.
is_retrieval_intent_template will be removed in Rasa 3.0.0. Please use is_retrieval_intent_response instead.
check_missing_templates will be removed in Rasa 3.0.0. Please use check_missing_responses instead.

Response Selector

The field template_name will be deprecated in Rasa 3.0.0. Please use utter_action instead. Please see here for more details.
The field response_templates will be deprecated in Rasa 3.0.0. Please use responses instead. Please see here for more details.

Rasa 2.3.3 to Rasa 2.3.4

caution

This is a release breaking backwards compatibility of machine learning models. It is not possible to load previously trained models if they were trained with model_confidence=cosine or model_confidence=inner setting. Please make sure to re-train the assistant before trying to use it with this improved version.

Machine Learning Components

Rasa 2.3.0 introduced the option of using cosine similarities for model confidences by setting model_confidence=cosine. Some post-release experiments revealed that using model_confidence=cosine is wrong as it can change the order of predicted labels. That's why this option was removed in Rasa version 2.3.4.

model_confidence=inner is deprecated as it produces an unbounded range of confidences which can break the logic of assistants in various other places.

We encourage you to try model_confidence=linear_norm which will produce a linearly normalized version of dot product similarities with each value in the range [0,1]. This can be done with the following config:

- name: DIETClassifier
  model_confidence: linear_norm
  constrain_similarities: True

If you trained a model with model_confidence=cosine or model_confidence=inner setting using previous versions of Rasa, please re-train by either removing the model_confidence option from the configuration or setting it to linear_norm.

Rasa 2.2 to Rasa 2.3

General

If you want to use Tensorboard for DIETClassifier, ResponseSelector, or TEDPolicy and log metrics after every (mini)batch, please use 'batch' instead of 'minibatch' as 'tensorboard_log_level'.

Machine Learning Components

A few changes have been made to the loss function inside machine learning (ML) components DIETClassifier, ResponseSelector and TEDPolicy. These include:

Configuration option loss_type=softmax is now deprecated and will be removed in Rasa 3.0.0. Use loss_type=cross_entropy instead.
The default loss function (loss_type=cross_entropy) can add an optional sigmoid cross-entropy loss of all similarity values to constrain them to an approximate range. You can turn on this option by setting constrain_similarities=True. This should help the models to perform better on real world test sets.

A new option model_confidence has been added to each ML component. It affects how the model's confidence for each label is computed during inference. It can take one of three values:

softmax - Dot product similarities between input and label embeddings are post-processed with a softmax function, as a result of which confidence for all labels sum up to 1.
cosine - Cosine similarity between input and label embeddings. Confidence for each label will be in the range [-1,1].
linear_norm - Dot product similarities between input and label embeddings are post-processed with a linear normalization function. Confidence for each label will be in the range [0,1].

The default value is softmax, but we recommend trying linear_norm. This should make it easier to tune thresholds for triggering fallback. The value of this option does not affect how confidences are computed for entity predictions in DIETClassifier.

We encourage you to try both the above recommendations. This can be done with the following config:

- name: DIETClassifier
  model_confidence: linear_norm
  constrain_similarities: True
  ...

Once the assistant is re-trained with the above configuration, users should also tune fallback confidence thresholds.

EDIT: Some post-release experiments revealed that using model_confidence=cosine is wrong as it can change the order of predicted labels. That's why this option was removed in Rasa version 2.3.4.

Rasa 2.1 to Rasa 2.2

General

TEDPolicy's transformer_size, number_of_transformer_layers, and dense_dimensions parameters have been renamed. Please update your configuration files using the following mapping:

Old Model Parameter	New Model Parameter
`transformer_size`	dictionary `transformer_size` with keys
	`text`, `action_text`, `label_action_text`, `dialogue`
`number_of_transformer_layers`	dictionary `number_of_transformer_layers` with keys
	`text`, `action_text`, `label_action_text`, `dialogue`
`dense_dimension`	dictionary `dense_dimension` with keys
	`text`, `action_text`, `label_action_text`, `intent`,
	`action_name`, `label_action_name`, `entities`, `slots`,
	`active_loop`

For example:

config.yml

policies:
  - name: TEDPolicy
    transformer_size:
      text: 128
      action_text: 128
      label_action_text: 128
      dialogue: 128
    number_of_transformer_layers:
      text: 1
      action_text: 1
      label_action_text: 1
      dialogue: 1
    dense_dimension:
      text: 128
      action_text: 128
      label_action_text: 128
      intent: 20
      action_name: 20
      label_action_name: 20
      entities: 20
      slots: 20
      active_loop: 20

Deprecations

Markdown Data

Training and test data in Markdown format is now deprecated. This includes:

reading and writing of story files in Markdown format
reading and writing of NLU data in Markdown format
reading and writing of retrieval intent data in Markdown format

Support for Markdown data will be removed entirely in Rasa 3.0.0.

Please convert your existing Markdown data by using the commands described here.

Policies

Policies now require a **kwargs argument in their constructor and load method. Policies without **kwargs will be supported until Rasa version 3.0.0. However when using incremental training **kwargs must be included.

Other

Domain.random_template_for is deprecated and will be removed in Rasa 3.0.0. You can alternatively use the TemplatedNaturalLanguageGenerator.
Domain.action_names is deprecated and will be removed in Rasa 3.0.0. Please use Domain.action_names_or_texts instead.

Rasa 2.0 to Rasa 2.1

Deprecations

ConveRTTokenizer is now deprecated. ConveRTFeaturizer now implements its behaviour. To migrate, replace ConveRTTokenizer with any other tokenizer, for e.g.:

pipeline:
    - name: WhitespaceTokenizer
    - name: ConveRTFeaturizer
      model_url: <Remote/Local path to model files>
    ...

HFTransformersNLP and LanguageModelTokenizer components are now deprecated. LanguageModelFeaturizer now implements their behaviour. To migrate, replace both the above components with any tokenizer and specify the model architecture and model weights as part of LanguageModelFeaturizer, for e.g.:

pipeline:
    - name: WhitespaceTokenizer
    - name: LanguageModelFeaturizer
      model_name: "bert"
      model_weights: "rasa/LaBSE"
    ...

Rasa 1.10 to Rasa 2.0

General

A lot has changed in version 2.0. Make sure you read through this guide thoroughly, to make sure all parts of your bot are updated. A lot of updates can be done automatically with inbuilt commands, others will need some manual conversion. If you have any feedback about these updates or the migration process, please post it in the forum.

Training data files

As of version 2.0, the new default training data format is yaml. Markdown is still supported, but this will be deprecated in Rasa 3.0.0.

You can convert existing NLU, Stories, and NLG (i.e. responses.md) training data files in the Markdown format to the new YAML format using following commands:

rasa data convert nlu -f yaml --data={SOURCE_DIR} --out={TARGET_DIR}
rasa data convert nlg -f yaml --data={SOURCE_DIR} --out={TARGET_DIR}
rasa data convert core -f yaml --data={SOURCE_DIR} --out={TARGET_DIR}

Converted files will have the same names as the original ones but with a _converted.yml suffix.

If you are using forms or response selectors, some additional changes will need to be made as described in their respective sections.

Policies

With the introduction of rules and the RulePolicy, the following policies are deprecated:

To migrate the policies automatically, you can run the following command:

rasa data convert config

This command will take care of updating your config.yml and domain.yml, while making backups of your existing files using the .bak suffix. It will also add a rules.yml if necessary.

Your forms will still function as normal in the old format after this update, but this command does not convert them into the new format automatically. This should be done manually, as described in the section on forms.

You can also migrate the individual policies manually, if you don't want to use the automatic conversion command.

Manually migrating from the Mapping Policy

If you previously used the Mapping Policy, you can follow the documentation on FAQs to convert your mapped intents to rules. Suppose you previously mapped an intent ask_is_bot as follows:

domain.yml

intents:
 - ask_is_bot:
     triggers: action_is_bot

This becomes the following rule:

rules.yml

rules:
- rule: Rule to map `ask_is_bot` intent
  steps:
  - intent: ask_is_bot
  - action: action_is_bot

And you can safely remove any triggers: from your domain:

domain.yml

intents:
 - ask_is_bot

Finally, you can replace the Mapping Policy with the Rule Policy in your model configuration:

config.yml

policies:
  # Other policies
  - name: RulePolicy

Manually migrating from the Fallback Policy

If you previously used the Fallback Policy, the following model configuration would translate as follows given a previous configuration like this:

config.yml

policies:
  - name: "FallbackPolicy"
    nlu_threshold: 0.4
    core_threshold: 0.3
    fallback_action_name: "action_default_fallback"
    ambiguity_threshold: 0.1

The new configuration would then look like:

config.yml

recipe: default.v1
pipeline:
  # Other components
  - name: FallbackClassifier
    threshold: 0.4
    ambiguity_threshold: 0.1

policies:
  # Other policies
  - name: RulePolicy
    core_fallback_threshold: 0.3
    core_fallback_action_name: "action_default_fallback"

In addition, you need to add a rule to specify which action to run in case of low NLU confidence:

rules.yml

rules:
  - rule: Ask the user to rephrase whenever they send a message with low NLU confidence
    steps:
    - intent: nlu_fallback
    - action: utter_please_rephrase

See the documentation on fallback for more information.

Manually migrating from the Two-Stage-Fallback Policy

If you previously used the Two-Stage Fallback Policy, with a configuration like this for example:

config.yml

policies:
  - name: TwoStageFallbackPolicy
    nlu_threshold: 0.4
    ambiguity_threshold: 0.1
    core_threshold: 0.3
    fallback_core_action_name: "action_default_fallback"
    fallback_nlu_action_name: "action_default_fallback"
    deny_suggestion_intent_name: "out_of_scope"

The new configuration would look like this:

config.yml

recipe: default.v1
pipeline:
  # Other components
  - name: FallbackClassifier
    threshold: 0.4
    ambiguity_threshold: 0.1

policies:
  # Other policies
  - name: RulePolicy
    core_fallback_threshold: 0.3
    core_fallback_action_name: "action_default_fallback"

In addition you need to add a rule to activate the Two-Stage Fallback for messages with low NLU confidence.

rules.yml

rules:
  - rule: Implementation of the TwoStageFallbackPolicy
    steps:
    # This intent is automatically triggered by the `FallbackClassifier` in the NLU
    # pipeline in case the intent confidence was below the specified threshold.
    - intent: nlu_fallback
    # The Fallback is now implemented as a form.
    - action: action_two_stage_fallback
    - active_loop: action_two_stage_fallback

Note that the previous parameters fallback_nlu_action_name and deny_suggestion_intent_name are no longer configurable and have the fixed values action_default_fallback and out_of_scope.

See the fallback documentation for more information.

Forms

As of version 2.0 the logic for forms has been moved from the Rasa SDK to Rasa to simplify implementation and make it easier to write action servers in other languages.

This means that forms are no longer implemented using a FormAction, but instead defined in the domain. Any customizations around requesting slots or slot validation can be handled with a FormValidationAction.

Consider a custom form action from 1.x like this:

from typing import Text, List, Any, Dict, Union
from rasa_sdk import Tracker
from rasa_sdk.executor import CollectingDispatcher
from rasa_sdk.forms  import FormAction

class RestaurantForm(FormAction):
    def name(self) -> Text:
        return "restaurant_form"

    @staticmethod
    def required_slots(tracker: Tracker) -> List[Text]:
        return ["cuisine"]

    def slot_mappings(self) -> Dict[Text, Union[Dict, List[Dict]]]:
        return {
            "cuisine": self.from_entity(entity="cuisine", not_intent="chitchat"),
        }

    @staticmethod
    def cuisine_db() -> List[Text]:
        """Database of supported cuisines"""

        return ["caribbean", "chinese", "french"]

    def validate_cuisine(
        self,
        value: Text,
        dispatcher: CollectingDispatcher,
        tracker: Tracker,
        domain: Dict[Text, Any],
    ) -> Dict[Text, Any]:
        """Validate cuisine value."""

        if value.lower() in self.cuisine_db():
            # validation succeeded, set the value of the "cuisine" slot to value
            return {"cuisine": value}
        else:
            dispatcher.utter_message(template="utter_wrong_cuisine")
            # validation failed, set this slot to None, meaning the
            # user will be asked for the slot again
            return {"cuisine": None}

    def submit(
        self,
        dispatcher: CollectingDispatcher,
        tracker: Tracker,
        domain: Dict[Text, Any],
    ) -> List[Dict]:
        """Define what the form has to do
            after all required slots are filled"""

        # utter submit template
        dispatcher.utter_message(template="utter_submit")
        return []

Start the migration by removing the FormPolicy and adding the RulePolicy (if not there already) to your model configuration:

config.yml

policies:
  # Other policies
  # ...
  - name: RulePolicy

Then you need to define the form, required slots and their slot mappings in the domain as described in the documentation on forms:

domain.yml

forms:
  restaurant_form:
    cuisine:
    - type: from_entity
      entity: cuisine
      not_intent: chitchat

If you ran the command to convert your stories, you will have a story that handles form activation and deactivation like this:

stories.yml

stories:
  - story: cuisine form
    steps:
    - intent: request_restaurant
    - action: restaurant_form
    - active_loop: restaurant_form
    - active_loop: null
    - action: utter_submit

This will work fine, but the best way to handle form behavior is to remove this story and instead define two separate rules for form activation and submission:

rules.yml

rules:
  - rule: Activate form
    steps:
    - intent: request_restaurant
    - action: restaurant_form
    - active_loop: restaurant_form

  - rule: Submit form
    condition:
    # Condition that form is active.
    - active_loop: restaurant_form
    steps:
    - action: restaurant_form
    - active_loop: null
    # The action we want to run when the form is submitted.
    - action: utter_submit

The last step is to implement a custom action to validate the form slots. Start by adding the custom action to your domain:

domain.yml

actions:
  # Other actions
  # ...
  - validate_restaurant_form

Then add a custom action which validates the cuisine slot:

from typing import Text, List, Any, Dict, Union
from rasa_sdk import Tracker
from rasa_sdk.executor import CollectingDispatcher
from rasa_sdk import FormValidationAction
from rasa_sdk.types import DomainDict

class RestaurantFormValidator(FormValidationAction):
    def name(self) -> Text:
        return "validate_restaurant_form"

    @staticmethod
    def cuisine_db() -> List[Text]:
        """Database of supported cuisines"""

        return ["caribbean", "chinese", "french"]

    def validate_cuisine(
        self,
        slot_value: Any,
        dispatcher: CollectingDispatcher,
        tracker: Tracker,
        domain: DomainDict,
    ) -> Dict[Text, Any]:
        """Validate cuisine value."""

        if slot_value.lower() in self.cuisine_db():
            # validation succeeded, set the value of the "cuisine" slot to value
            return {"cuisine": slot_value}
        else:
            # validation failed, set this slot to None, meaning the
            # user will be asked for the slot again
            return {"cuisine": None}

You can also migrate forms from Rasa SDK to Rasa 2 iteratively. You can for example migrate one form to the Rasa 2 implementation while continue using the deprecated Rasa SDK implementation for another form. To continue to use the deprecated Rasa SDK FormActions, add a custom action with the name of your form to your domain. Note that you should complete the migration as soon as possible as the deprecated FormAction will be removed from the Rasa SDK in Rasa 3.

domain.yml

actions:
# Adding a custom action for a form will
# instruct Rasa to use the
# deprecated Rasa SDK implementation of forms.
- my_form

forms:
 my_form:

See the forms documentation for more details.

Response Selectors

Response Selectors are a stable feature as of version 2.0.

The conversion command will automatically convert your responses.md file, stories and nlu training data to the new yaml format. It will also take care of adding the utter_ prefix to your responses. Additionally you will need to rename the respond_ actions in your stories files to use the utter_ prefix instead. Run the following command to apply these changes:

rasa data convert responses --data {SOURCE_DIR} --out={TARGET_DIR}

You can also apply these changes manually. For example:

stories:
  - story: chitchat
    steps:
    - intent: chitchat
    - action: respond_chitchat

becomes

stories:
  - story: chitchat
    steps:
    - intent: chitchat
    - action: utter_chitchat

and you will need to add the utter_ prefix to the response names in your responses.md as well. For example:

responses:
  chitchat/ask_name:
    - text: Oh yeah, I am called the retrieval bot.

  chitchat/ask_weather:
    - text: Oh, it does look sunny right now in Berlin.

becomes

responses:
  utter_chitchat/ask_name:
    - text: Oh yeah, I am called the retrieval bot.

  utter_chitchat/ask_weather:
    - text: Oh, it does look sunny right now in Berlin.

Finally, you should remove any actions with the respond_ prefix from the actions list in your domain.

This behavior will work fine when defined as a story, but even better when defined as a rule. You should consider transferring your retrieval stories to rules. More information on what that looks like in the chitchat and FAQs documentation.

Response Selectors are now trained on retrieval intent labels by default instead of the actual response text. For most models, this should improve training time and accuracy of the ResponseSelector.

If you want to revert to the pre-2.0 default behavior, add the use_text_as_label: true parameter to your ResponseSelector component:

pipeline:
  # other components
  - name: ResponseSelector
    use_text_as_label: true

The output schema of ResponseSelector has changed. An example output looks like this:

{
  "response_selector": {
    "all_retrieval_intents": [
      "faq"
    ],
    "default": {
      "response": {
        "id": 1388783286124362000,
        "confidence": 1,
        "intent_response_key": "faq/is_legit",
        "response_templates": [
          {
            "text": "absolutely",
            "image": "https://i.imgur.com/nGF1K8f.jpg"
          },
          {
            "text": "I think so."
          }
        ]
        "template_name": "utter_faq/is_legit"
      },
      "ranking": [
        {
          "id": 1388783286124362000,
          "confidence": 1,
          "intent_response_key": "faq/is_legit"
        }
      ]
    }
  }
}

As a result of this, if you were previously querying for the key full_retrieval_intent as:

response_selector_output.get("default")
                        .get("full_retrieval_intent")

you should instead now do this:

response_selector_output.get("default")
                        .get("response")
                        .get("intent_response_key")

Unfeaturized Slots

Slots of type unfeaturized are deprecated and will be removed in version 3.0. To ignore slot values during a conversation, set the influence_conversation property of the slot to false.

The following snippet is an example of the deprecated unfeaturized slot usage:

slots:
  username:
    type: unfeaturized

To update this to the new format, you can specify the expected data type text and define that the slot should be ignored during the conversation.

slots:
  username:
    type: text
    # Set `influence_conversation` to `false`
    # to ignore the slot value during the conversation.
    influence_conversation: false

If you don't require the slot to have a specific data type, you can use the new slot type any. This slot type is always ignored during a conversation and does not make any assumptions regarding the data type of the slot value.

slots:
  username:
    type: any

Please see the updated slots documentation for more information.

Conversation sessions

Conversation sessions are now enabled by default if your Domain does not contain a session configuration. Previously a missing session configuration was treated as if conversation sessions were disabled. You can explicitly disable conversation sessions using the following snippet:

domain.yml

session_config:
  # A session expiration time of `0`
  # disables conversation sessions
  session_expiration_time: 0

Dialogue Featurization

This section is only relevant if you explicitly defined featurizers in your policy configuration.

LabelTokenizerSingleStateFeaturizer is deprecated and will be removed in the future. It should be replaced with SingleStateFeaturizer and some changes should be made to the NLU pipeline. Add a Tokenizer with the option intent_tokenization_flag: True and CountVectorsFeaturizer to the NLU pipeline.

For example:

language: en
pipeline:
  - name: WhitespaceTokenizer
    intent_tokenization_flag: True
  - name: CountVectorsFeaturizer
  # other components
policies:
  # other policies
  - name: TEDPolicy
    featurizer:
    - name: SingleStateFeaturizer

BinarySingleStateFeaturizer is deprecated and will be removed in the future. You should replace it with SingleStateFeaturizer and a NLU pipeline where intent_tokenization_flag of a Tokenizer is set to False.

For example:

language: en
pipeline:
  - name: WhitespaceTokenizer
    intent_tokenization_flag: False
  # other components
policies:
  # other policies
  - name: TEDPolicy
    featurizer:
    - name: SingleStateFeaturizer

Deprecations

The deprecated event brokers FileProducer, KafkaProducer, PikaProducer and SQLProducer have been removed. If you used these brokers in your endpoints.yml make sure to use the renamed variants instead:

FileProducer became FileEventBroker
KafkaProducer became KafkaEventBroker
PikaProducer became PikaEventBroker
SQLProducer became SQLEventBroker

The deprecated EmbeddingIntentClassifier has been removed. If you used this component in your pipeline configuration (config.yml) you can replace it with DIETClassifier. It accepts the same configuration parameters.

The deprecated KerasPolicy has been removed. If you used this component in your policies configuration (config.yml) you can replace it with TEDPolicy. It accepts the same configuration parameters.

Rasa 1.7 to Rasa 1.8

caution

This is a release breaking backwards compatibility. It is not possible to load previously trained models. Please make sure to retrain a model before trying to use it with this improved version.

General

The TED Policy replaced the keras_policy as recommended machine learning policy. New projects generated with rasa init will automatically use this policy. In case you want to change your existing model configuration to use the TED Policy add this to the policies section in your config.yml and remove potentially existing KerasPolicy entries:
policies:
# - ... other policies
- name: TEDPolicy
max_history: 5
epochs: 100
The given snippet specifies default values for the parameters max_history and epochs. max_history is particularly important and strongly depends on your stories. Please see the docs of the TED Policy if you want to customize them.
All pre-defined pipeline templates are deprecated. Any templates you use will be mapped to the new configuration, but the underlying architecture is the same. Take a look at Tuning Your Model to decide on what components you should use in your configuration file.
The Embedding Policy was renamed to TED Policy. The functionality of the policy stayed the same. Please update your configuration files to use TEDPolicy instead of EmbeddingPolicy.

Most of the model options for EmbeddingPolicy, EmbeddingIntentClassifier, and ResponseSelector got renamed. Please update your configuration files using the following mapping:

Old model option	New model option
hidden_layers_sizes_a	dictionary “hidden_layers_sizes” with key “text”
hidden_layers_sizes_b	dictionary “hidden_layers_sizes” with key “label”
hidden_layers_sizes_pre_dial	dictionary “hidden_layers_sizes” with key “dialogue”
hidden_layers_sizes_bot	dictionary “hidden_layers_sizes” with key “label”
num_transformer_layers	number_of_transformer_layers
num_heads	number_of_attention_heads
max_seq_length	maximum_sequence_length
dense_dim	dense_dimension
embed_dim	embedding_dimension
num_neg	number_of_negative_examples
mu_pos	maximum_positive_similarity
mu_neg	maximum_negative_similarity
use_max_sim_neg	use_maximum_negative_similarity
C2	regularization_constant
C_emb	negative_margin_scale
droprate_a	droprate_dialogue
droprate_b	droprate_label
evaluate_every_num_epochs	evaluate_every_number_of_epochs
evaluate_on_num_examples	evaluate_on_number_of_examples

Old configuration options will be mapped to the new names, and a warning will be thrown. However, these will be deprecated in a future release.

The Embedding Intent Classifier is now deprecated and will be replaced by DIETClassifier in the future. DIETClassfier performs intent classification as well as entity recognition. If you want to get the same model behavior as the current EmbeddingIntentClassifier, you can use the following configuration of DIETClassifier:
pipeline:
# - ... other components
- name: DIETClassifier
hidden_layers_sizes:
text: [256, 128]
number_of_transformer_layers: 0
weight_sparsity: 0
intent_classification: True
entity_recognition: False
use_masked_language_model: False
BILOU_flag: False
scale_loss: True
use_sparse_input_dropout: False
use_dense_input_dropout: False
# ... any other parameters
See DIETClassifier for more information about the new component. Specifying EmbeddingIntentClassifier in the configuration maps to the above component definition, and results in the same behaviour within the same Rasa version.
CRFEntityExtractor is now deprecated and will be replaced by DIETClassifier in the future. If you want to get the same model behavior as the current CRFEntityExtractor, you can use the following configuration:
pipeline:
# - ... other components
- name: LexicalSyntacticFeaturizer
features: [
["low", "title", "upper"],
[
"BOS",
"EOS",
"low",
"prefix5",
"prefix2",
"suffix5",
"suffix3",
"suffix2",
"upper",
"title",
"digit",
],
["low", "title", "upper"],
]
- name: DIETClassifier
intent_classification: False
entity_recognition: True
use_masked_language_model: False
number_of_transformer_layers: 0
# ... any other parameters
CRFEntityExtractor featurizes user messages on its own, it does not depend on any featurizer. We extracted the featurization from the component into the new featurizer LexicalSyntacticFeaturizer. Thus, in order to obtain the same results as before, you need to add this featurizer to your pipeline before the DIETClassifier. Specifying CRFEntityExtractor in the configuration maps to the above component definition, the behavior is unchanged from previous versions.
If your pipeline contains CRFEntityExtractor and EmbeddingIntentClassifier you can substitute both components with DIETClassifier. You can use the following pipeline for that:
pipeline:
# - ... other components
- name: LexicalSyntacticFeaturizer
features: [
["low", "title", "upper"],
[
"BOS",
"EOS",
"low",
"prefix5",
"prefix2",
"suffix5",
"suffix3",
"suffix2",
"upper",
"title",
"digit",
],
["low", "title", "upper"],
]
- name: DIETClassifier
number_of_transformer_layers: 0
# ... any other parameters

Rasa 1.6 to Rasa 1.7

General

By default, the EmbeddingIntentClassifier, EmbeddingPolicy, and ResponseSelector will now normalize the top 10 confidence results if the loss_type is "softmax" (which has been default since 1.3, see Rasa 1.2 to Rasa 1.3). This is configurable via the ranking_length configuration parameter; to turn off normalization to match the previous behavior, set ranking_length: 0.

Rasa 1.2 to Rasa 1.3

caution

This is a release breaking backwards compatibility. It is not possible to load previously trained models. Please make sure to retrain a model before trying to use it with this improved version.

General

Default parameters of EmbeddingIntentClassifier are changed. See the Components page for details. Architecture implementation is changed as well, so old trained models cannot be loaded. Default parameters and architecture for EmbeddingPolicy are changed. See Policies for details. It uses transformer instead of lstm. Old trained models cannot be loaded. They use inner similarity and softmax loss by default instead of cosine similarity and margin loss (can be set in config file). They use balanced batching strategy by default to counteract class imbalance problem. The meaning of evaluate_on_num_examples is changed. If it is non zero, random examples will be picked by stratified split and used as hold out validation set, so they will be excluded from training data. We suggest to set it to zero (default) if data set contains a lot of unique examples of dialogue turns. Removed label_tokenization_flag and label_split_symbol from component. Instead moved intent splitting to Tokenizer components via intent_tokenization_flag and intent_split_symbol flag.
Default max_history for EmbeddingPolicy is None which means it'll use the FullDialogueTrackerFeaturizer. We recommend to set max_history to some finite value in order to use MaxHistoryTrackerFeaturizer for faster training. See Featurizers for details. We recommend to increase batch_size for MaxHistoryTrackerFeaturizer (e.g. "batch_size": [32, 64])
Compare mode of rasa train core allows the whole core config comparison. Therefore, we changed the naming of trained models. They are named by config file name instead of policy name. Old naming style will not be read correctly when creating compare plots (rasa test core). Please remove old trained models in comparison folder and retrain. Normal core training is unaffected.
We updated the evaluation metric for our NER. We report the weighted precision and f1-score. So far we included no-entity in this report. However, as most of the tokens actually don't have an entity set, this will influence the weighted precision and f1-score quite a bit. From now on we exclude no-entity from the evaluation. The overall metrics now only include proper entities. You might see a drop in the performance scores when running the evaluation again.
/ is reserved as a delimiter token to distinguish between retrieval intent and the corresponding response text identifier. Make sure you don't include / symbol in the name of your intents.

Rasa NLU 0.14.x and Rasa Core 0.13.x to Rasa 1.0

caution

This is a release breaking backwards compatibility. It is not possible to load previously trained models. Please make sure to retrain a model before trying to use it with this improved version.

General

The scripts in rasa.core and rasa.nlu can no longer be executed. To train, test, run, … an NLU or Core model, you should now use the command line interface rasa. The functionality is, for the most part, the same as before. Some changes in commands reflect the combined training and running of NLU and Core models, but NLU and Core can still be trained and used individually. If you attempt to run one of the old scripts in rasa.core or rasa.nlu, an error is thrown that points you to the command you should use instead. See all the new commands at Command Line Interface.
If you have written a custom output channel, all send_ methods subclassed from the OutputChannel class need to take an additional \*\*kwargs argument. You can use these keyword args from your custom action code or the templates in your domain file to send any extra parameters used in your channel's send methods.
If you were previously importing the Button or Element classes from rasa_core.dispatcher, these are now to be imported from rasa_sdk.utils.
Rasa NLU and Core previously used separate configuration files. These two files should be merged into a single file either named config.yml, or passed via the --config parameter.

Script parameters

All script parameter names have been unified to follow the same schema. Any underscores (_) in arguments have been replaced with dashes (-). For example: --max_history has been changed to --max-history. You can see all of the script parameters in the --help output of the commands in the Command Line Interface.
The --num_threads parameter was removed from the run command. The server will always run single-threaded, but will now run asynchronously. If you want to make use of multiple processes, feel free to check out the Sanic server documentation.
To avoid conflicts in script parameter names, connectors in the run command now need to be specified with --connector, as -c is no longer supported. The maximum history in the rasa visualize command needs to be defined with --max-history. Output paths and log files cannot be specified with -o anymore; --out and --log-file should be used. NLU data has been standarized to be --nlu and the name of any kind of data files or directory to be --data.

REST API

There are numerous REST API endpoint changes which can be found here.

Rasa Pro 3.9 to Rasa Pro 3.10#

LLM/Embedding Configuration#

note

Disabling the cache#

Custom Components using an LLM#

Changes to llm_factory#

Changes to embedder_factory#

Changes to invoke_llm#

Changes to SingleStepLLMCommandGenerator.predict_commands#

Changes to the default configuration dictionary#

Dropped support for Python 3.8#

Rasa Pro 3.8 to Rasa Pro 3.9#

LLMCommandGenerator#

Slot Mappings#

Rasa Pro 3.8.0 to Rasa Pro 3.8.1#

Poetry Installation#

Rasa Pro 3.7 to 3.8#

info

Installation#

Python package#

Helm Chart / Docker Image#

Component Yaml Configuration Changes#

Changes to default behaviour#

info

Prompt Rendering#

Processing Chitchat#

Handling of categorical slots#

Update default signature of LLM calls#

Dependency Upgrades#

Spacy 3.7.x#

Pydantic 2.x#

Rasa Pro 3.7.9 to Rasa Pro 3.7.10#

Poetry Installation#

Rasa Pro 3.7.8 to Rasa Pro 3.7.9#

Changes to default behaviour#

Handling of categorical slots#

Rasa 3.6 to Rasa Pro 3.7#

Installation#

info

Python package#

Helm Chart / Docker Image#

Rasa 3.0 to 3.1#

Machine Learning Components#

TensorFlow Upgrade#

NLU JSON Format#

Rasa 2.x to 3.0#

Markdown Data#

Model Configuration#

Custom Policies and Custom Components#

caution

Forward References with Python 3.7

Changes to Custom NLU Components#

Changes to Custom Policies#

caution

Training data#

Upgrading version from 2.0 to 3.0#

TrainingDataImporter#

Training#

rasa train --dry-run#

Machine Learning Components#

Normalization of Confidences in DIETClassifier and ResponseSelector#

Normalization of confidences in TEDPolicy#

Removed Policies#

Removed Tokenizers and Featurizers#

Slot Mappings#

note

Automatic migration from 2.0 domain format to the 3.0 format#

Manually migrating from 2.0 domain format to the 3.0 format#

Rasa-SDK Modifications#

note

Rasa 2.7 to 2.8#

caution

Deprecations#

Tracker Featurizers#

State Featurizer#

Incremental Training#

Machine Learning Components#

Rasa 2.5 to 2.6#

Forms#

New ignored_intents parameter in Forms#

Rasa Pro 3.9 to Rasa Pro 3.10

LLM/Embedding Configuration

Disabling the cache

Custom Components using an LLM

Changes to `llm_factory`

Changes to `embedder_factory`

Changes to `invoke_llm`

Changes to `SingleStepLLMCommandGenerator.predict_commands`

Changes to the default configuration dictionary

Dropped support for Python 3.8

Rasa Pro 3.8 to Rasa Pro 3.9

LLMCommandGenerator

Slot Mappings

Rasa Pro 3.8.0 to Rasa Pro 3.8.1

Poetry Installation

Rasa Pro 3.7 to 3.8

Installation

Python package

Helm Chart / Docker Image

Component Yaml Configuration Changes

Changes to default behaviour

Prompt Rendering

Processing Chitchat

Handling of categorical slots

Update default signature of LLM calls

Dependency Upgrades

Spacy 3.7.x

Pydantic 2.x

Rasa Pro 3.7.9 to Rasa Pro 3.7.10

Poetry Installation

Rasa Pro 3.7.8 to Rasa Pro 3.7.9

Changes to default behaviour

Handling of categorical slots

Rasa 3.6 to Rasa Pro 3.7

Installation

Python package

Helm Chart / Docker Image

Rasa 3.0 to 3.1

Machine Learning Components

TensorFlow Upgrade

NLU JSON Format

Rasa 2.x to 3.0

Markdown Data

Model Configuration

Custom Policies and Custom Components

Changes to Custom NLU Components

Changes to Custom Policies

Training data

Upgrading `version` from `2.0` to `3.0`

`TrainingDataImporter`

Training

`rasa train --dry-run`

Machine Learning Components

Normalization of Confidences in `DIETClassifier` and `ResponseSelector`

Normalization of confidences in `TEDPolicy`

Removed Policies

Removed Tokenizers and Featurizers

Slot Mappings

Automatic migration from 2.0 domain format to the 3.0 format

Manually migrating from 2.0 domain format to the 3.0 format

Rasa-SDK Modifications

Rasa 2.7 to 2.8

Deprecations

Tracker Featurizers

State Featurizer

Incremental Training

Machine Learning Components

Rasa 2.5 to 2.6

Forms

New `ignored_intents` parameter in Forms

Rasa 2.4 to 2.5

Machine Learning Components

`DIET`, `TED`, and `ResponseSelector`

SpaCy 3.0

Rasa 2.3 to Rasa 2.4

Deprecating `template` for `response`

Rasa 2.3.3 to Rasa 2.3.4

Machine Learning Components

Rasa 2.2 to Rasa 2.3

General