Release 2.0.0 · RasaHQ/rasa

Deprecations and Removals

#5757: Removed previously deprecated packages rasa_nlu and rasa_core.

Use imports from rasa.core and rasa.nlu instead.
#5758: Removed previously deprecated classes:
- event brokers (EventChannel and FileProducer, KafkaProducer,
  PikaProducer, SQLProducer)
- intent classifier EmbeddingIntentClassifier
- policy KerasPolicy
Removed previously deprecated methods:
- Agent.handle_channels
- TrackerStore.create_tracker_store
Removed support for pipeline templates in config.yml

Removed deprecated training data keys entity_examples and intent_examples from
json training data format.
#5834: Removed restaurantbot example as it was confusing and not a great way to build a bot.
#6296: LabelTokenizerSingleStateFeaturizer is deprecated. To replicate LabelTokenizerSingleStateFeaturizer functionality,
add a Tokenizer with intent_tokenization_flag: True and CountVectorsFeaturizer to the NLU pipeline.
An example of elements to be added to the pipeline is shown in the improvement changelog 6296`.

BinarySingleStateFeaturizer is deprecated and will be removed in the future. We recommend to switch to SingleStateFeaturizer.
#6354: Specifying the parameters force and save_to_default_model_directory as part of the
JSON payload when training a model using POST /model/train is now deprecated.
Please use the query parameters force_training and save_to_default_model_directory
instead. See the API documentation for more information.
#6409: The conversation event form was renamed to active_loop. Rasa Open Source
will continue to be able to read and process old form events. Note that
serialized trackers will no longer have the active_form field. Instead the
active_loop field will contain the same information. Story representations
in Markdown and YAML will use active_loop instead of form to represent the
event.
#6453: Removed support for queue argument in PikaEventBroker (use queues instead).

Domain file:
- Removed support for templates key (use responses instead).
- Removed support for string responses (use dictionaries instead).
NLU Component:
- Removed support for provides attribute, it's not needed anymore.
- Removed support for requires attribute (use required_components() instead).
Removed _guess_format() utils method from rasa.nlu.training_data.loading (use guess_format instead).

Removed several config options for TED Policy, DIETClassifier and ResponseSelector:
- hidden_layers_sizes_pre_dial
- hidden_layers_sizes_bot
- droprate
- droprate_a
- droprate_b
- hidden_layers_sizes_a
- hidden_layers_sizes_b
- num_transformer_layers
- num_heads
- dense_dim
- embed_dim
- num_neg
- mu_pos
- mu_neg
- use_max_sim_neg
- C2
- C_emb
- evaluate_every_num_epochs
- evaluate_on_num_examples
Please check the documentation for more information.
#6463: The conversation event form_validation was renamed to loop_interrupted.
Rasa Open Source will continue to be able to read and process old form_validation
events.
#6658: SklearnPolicy was deprecated. TEDPolicy is the preferred machine-learning policy for dialogue models.
#6809: Slots of type unfeaturized are
now deprecated and will be removed in Rasa Open Source 3.0. Instead you should use
the property influence_conversation: false for every slot type as described in the
migration guide.
#6934: 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:
```
session_config:
  # A session expiration time of `0`
  # disables conversation sessions
  session_expiration_time: 0
```
#6952: Using the default action action_deactivate_form to deactivate
the currently active loop / Form is deprecated.
Please use action_deactivate_loop instead.

Features

#4745: Added template name to the metadata of bot utterance events.

BotUttered event contains a template_name property in its metadata for any
new bot message.
#5086: Added a --num-threads CLI argument that can be passed to rasa train
and will be used to train NLU components.
#5510: You can now define what kind of features should be used by what component
(see Choosing a Pipeline).

You can set an alias via the option alias for every featurizer in your pipeline.
The alias can be anything, by default it is set to the full featurizer class name.
You can then specify, for example, on the
DIETClassifier what features from which
featurizers should be used.
If you don't set the option featurizers all available features will be used.
This is also the default behavior.
Check components to see what components have the option
featurizers available.

Here is an example pipeline that shows the new option.
We define an alias for all featurizers in the pipeline.
All features will be used in the DIETClassifier.
However, the ResponseSelector only takes the features from the
ConveRTFeaturizer and the CountVectorsFeaturizer (word level).
```
pipeline:
- name: ConveRTTokenizer
- name: ConveRTFeaturizer
  alias: "convert"
- name: CountVectorsFeaturizer
  alias: "cvf_word"
- name: CountVectorsFeaturizer
  alias: "cvf_char"
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 4
- name: RegexFeaturizer
  alias: "regex"
- name: LexicalSyntacticFeaturizer
  alias: "lsf"
- name: DIETClassifier:
- name: ResponseSelector
  epochs: 50
  featurizers: ["convert", "cvf_word"]
- name: EntitySynonymMapper
```
:::caution
This change is model-breaking. Please retrain your models.

:::
#5837: Added --port commandline argument to the interactive learning mode to allow
changing the port for the Rasa server running in the background.
#5957: Add new entity extractor RegexEntityExtractor. The entity extractor extracts entities using the lookup tables
and regexes defined in the training data. For more information see RegexEntityExtractor.
#5996: Introduced a new YAML format for Core training data and implemented a parser
for it. Rasa Open Source can now read stories in both Markdown and YAML format.
#6020: You can now enable threaded message responses from Rasa through the Slack connector.
This option is enabled using an optional configuration in the credentials.yml file
```
    slack:
      slack_token:
      slack_channel:
      use_threads: True
```
Button support has also been added in the Slack connector.
#6065: Add support for rules data and forms in YAML
format.
#6066: The NLU interpreter is now passed to the Policies during training and
inference time. Note that this requires an additional parameter interpreter in the
method predict_action_probabilities of the Policy interface. In case a
custom Policy implementation doesn't provide this parameter Rasa Open Source
will print a warning and omit passing the interpreter.
#6088: Added the new dialogue policy RulePolicy which will replace the old “rule-like”
policies Mapping Policy,
Fallback Policy,
Two-Stage Fallback Policy, and
Form Policy. These policies are now
deprecated and will be removed in the future. Please see the
rules documentation for more information.

Added new NLU component FallbackClassifier
which predicts an intent nlu_fallback in case the confidence was below a given
threshold. The intent nlu_fallback may
then be used to write stories / rules to handle the fallback in case of low NLU
confidence.
```
pipeline:
- # Other NLU components ...
- name: FallbackClassifier
  # If the highest ranked intent has a confidence lower than the threshold then
  # the NLU pipeline predicts an intent `nlu_fallback` which you can then be used in
  # stories / rules to implement an appropriate fallback.
  threshold: 0.5
```
#6132: Added possibility to split the domain into separate files. All YAML files
under the path specified with --domain will be scanned for domain
information (e.g. intents, actions, etc) and then combined into a single domain.

The default value for --domain is still domain.yml.
#6275: Add optional metadata argument to NaturalLanguageInterpreter's parse method.
#6354: The Rasa Open Source API endpoint POST /model/train now supports training data in YAML
format. Please specify the header Content-Type: application/yaml when
training a model using YAML training data.
See the API documentation for more information.
#6374: Added a YAML schema and a writer for 2.0 Training Core data.
#6404: Users can now use the rasa data convert {nlu|core} -f yaml command to convert training data from Markdown format to YAML format.
#6536: Add option use_lemma to CountVectorsFeaturizer. By default it is set to True.

use_lemma indicates whether the featurizer should use the lemma of a word for counting (if available) or not.
If this option is set to False it will use the word as it is.

Improvements

#4536: Add support for Python 3.8.
#5368: Changed the project structure for Rasa projects initialized with the
CLI (using the rasa init command):
actions.py -> actions/actions.py. actions is now a Python package (it contains
a file actions/__init__.py). In addition, the __init__.py at the
root of the project has been removed.
#5481: DIETClassifier now also assigns a confidence value to entity predictions.
#5637: Added behavior to the rasa --version command. It will now also list information
about the operating system, python version and rasa-sdk. This will make it easier
for users to file bug reports.
#5743: Support for additional training metadata.

Training data messages now to support kwargs and the Rasa JSON data reader
includes all fields when instantiating a training data instance.
#5748: Standardize testing output. The following test output can be produced for intents,
responses, entities and stories:
- report: a detailed report with testing metrics per label (e.g. precision,
  recall, accuracy, etc.)
- errors: a file that contains incorrect predictions
- successes: a file that contains correct predictions
- confusion matrix: plot of confusion matrix
- histogram: plot of confidence distribution (not available for stories)
#5756: To avoid the problem of our entity extractors predicting entity labels for
just a part of the words, we introduced a cleaning method after the prediction
was done. We should avoid the incorrect prediction in the first place.
To achieve this we will not tokenize words into sub-words anymore.
We take the mean feature vectors of the sub-words as the feature vector of the word.

:::caution
This change is model breaking. Please, retrain your models.

:::
#5759: Move option case_sensitive from the tokenizers to the featurizers.
- Remove the option from the WhitespaceTokenizer and ConveRTTokenizer.
- Add option case_sensitive to the RegexFeaturizer.
#5766: If a user sends a voice message to the bot using Facebook, users messages was set to the attachments URL. The same is now also done for the rest of attachment types (image, video, and file).
#5794: Creating a Domain using Domain.fromDict can no longer alter the input dictionary.
Previously, there could be problems when the input dictionary was re-used for other
things after creating the Domain from it.
#5805: The debug-level logs when instantiating an
SQLTrackerStore
no longer show the password in plain text. Now, the URL is displayed with the password
hidden, e.g. postgresql://username:***@localhost:5432.
#5855: Shorten the information in tqdm during training ML algorithms based on the log
level. If you train your model in debug mode, all available metrics will be
shown during training, otherwise, the information is shorten.
#5913: Ignore conversation test directory tests/ when importing a project
using MultiProjectImporter and use_e2e is False.
Previously, any story data found in a project subdirectory would be imported
as training data.
#5985: Implemented model checkpointing for DIET (including the response selector) and TED. The best model during training will be stored instead of just the last model. The model is evaluated on the basis of evaluate_every_number_of_epochs and evaluate_on_number_of_examples.

Checkpointing is enabled iff the following is set for the models in the config.yml file:
- checkpoint_model: True
- evaluate_on_number_of_examples > 0
The model is stored to whatever location has been specified with the --out parameter when calling rasa train nlu/core ....
#6024: rasa data split nlu now makes sure that there is at least one example per
intent and response in the test data.
#6039: The method ensure_consistent_bilou_tagging now also considers the confidence values of the predicted tags
when updating the BILOU tags.
#6045: We updated the way how we save and use features in our NLU pipeline.

The message object now has a dedicated field, called features, to store the
features that are generated in the NLU pipeline. We adapted all our featurizers in a
way that sequence and sentence features are stored independently. This allows us to
keep different kind of features for the sequence and the sentence. For example, the
LexicalSyntacticFeaturizer does not produce any sentence features anymore as our
experiments showed that those did not bring any performance gain just quite a lot of
additional values to store.

We also modified the DIET architecture to process the sequence and sentence
features independently at first. The features are concatenated just before
the transformer.

We also removed the __CLS__ token again. Our Tokenizers will not
add this token anymore.

:::caution
This change is model-breaking. Please retrain your models.

:::
#6052: Add endpoint kwarg to rasa.jupyter.chat to enable using a custom action server while chatting with a model in a jupyter notebook.
#6055: Support for rasa conversation id with special characters on the server side - necessary for some channels (e.g. Viber)
#6123: Add support for proxy use in slack input channel.
#6134: Log the number of examples per intent during training. Logging can be enabled using rasa train --debug.
#6237: Support for other remote storages can be achieved by using an external library.
#6273: Add output_channel query param to /conversations/<conversation_id>/tracker/events route, along with boolean execute_side_effects to optionally schedule/cancel reminders, and forward bot messages to output channel.
#6276: Allow Rasa to boot when model loading exception occurs. Forward HTTP Error responses to standard log output.
#6294: Rename DucklingHTTPExtractor to DucklingEntityExtractor.
#6296: * Modified functionality of SingleStateFeaturizer.

SingleStateFeaturizer uses trained NLU Interpreter to featurize intents and action names.
This modified SingleStateFeaturizer can replicate LabelTokenizerSingleStateFeaturizer functionality.
This component is deprecated from now on.
To replicate LabelTokenizerSingleStateFeaturizer functionality,
add a Tokenizer with intent_tokenization_flag: True and CountVectorsFeaturizer to the NLU pipeline.
Please update your configuration file.

For example:
yaml language: en pipeline: - name: WhitespaceTokenizer intent_tokenization_flag: True - name: CountVectorsFeaturizer

Please train both NLU and Core (using rasa train) to use a trained tokenizer and featurizer for core featurization.

The new SingleStateFeaturizer stores slots, entities and forms in sparse features for more lightweight storage.

BinarySingleStateFeaturizer is deprecated and will be removed in the future.
We recommend to switch to SingleStateFeaturizer.
- Modified TEDPolicy to handle sparse features. As a result, TEDPolicy may require more epochs than before to converge.
- Default TEDPolicy featurizer changed to MaxHistoryTrackerFeaturizer with infinite max history (takes all dialogue turns into account).
- Default batch size for TED increased from [8,32] to [64, 256]
#6323: Response selector templates now support all features that
domain utterances do. They use the yaml format instead of markdown now.
This means you can now use buttons, images, ... in your FAQ or chitchat responses
(assuming they are using the response selector).

As a consequence, training data form in markdown has to have the file
suffix .md from now on to allow proper file type detection-
#6457: Support for test stories written in yaml format.
#6466: 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.

You can now also have multiple response templates for a single sub-intent of a retrieval intent. The first response template
containing the text attribute is picked for training(if use_text_as_label=True) and a random template is picked for bot's utterance just as how other utter_ templates are picked.

All response selector related evaluation artifacts - report.json, successes.json, errors.json, confusion_matrix.png now use the sub-intent of the retrieval intent as the target and predicted labels instead of the actual response text.

The output schema of ResponseSelector has changed - full_retrieval_intent and name have been deprecated in favour
of intent_response_key and response_templates respectively. Additionally a key all_retrieval_intents
is added to the response selector output which will hold a list of all retrieval intents(faq,chitchat, etc.)
that are present in the training data.An example output looks like this -
```
"response_selector": {
    "all_retrieval_intents": ["faq"],
    "default": {
      "response": {
        "id": 1388783286124361986, "confidence": 1.0, "intent_response_key": "faq/is_legit",
        "response_templates": [
          {
            "text": "absolutely",
            "image": "https://i.imgur.com/nGF1K8f.jpg"
          },
          {
            "text": "I think so."
          }
        ],
      },
      "ranking": [
        {
          "id": 1388783286124361986,
          "confidence": 1.0,
          "intent_response_key": "faq/is_legit"
        },
      ]
```
An example bot demonstrating how to use the ResponseSelector is added to the examples folder.
#6472: Do not modify conversation tracker's latest_input_channel property when using POST /trigger_intent or ReminderScheduled.
#6555: Do not set the output dimension of the sparse-to-dense layers to the same dimension as the dense features.

Update default value of dense_dimension and concat_dimension for text in DIETClassifier to 128.
#6591: Retrieval actions with respond_ prefix are now replaced with usual utterance actions with utter_ prefix.

If you were using retrieval actions before, rename all of them to start with utter_ prefix. For example, respond_chitchat becomes utter_chitchat.
Also, in order to keep the response templates more consistent, you should now add the utter_ prefix to all response templates defined for retrieval intents. For example, a response template chitchat/ask_name becomes utter_chitchat/ask_name. Note that the NLU examples for this will still be under chitchat/ask_name intent.
The example responseselectorbot should help clarify these changes further.
#6613: Added telemetry reporting. Rasa uses telemetry to report anonymous usage information.
This information is essential to help improve Rasa Open Source for all users.
Reporting will be opt-out. More information can be found in our
telemetry documentation.
#6757: Update extract_other_slots method inside FormAction to fill a slot from an entity
with a different name if corresponding slot mapping of from_entity type is unique.
#6809: Slots of any type can now be ignored during a conversation.
To do so, specify the property influence_conversation: false for the slot.
```
slot:
  a_slot:
    type: text
    influence_conversation: false
```
The property influence_conversation is set to true by default. See the
documentation for slots for more information.

A new slot type any was added. Slots of this type can store
any value. Slots of type any are always ignored during conversations.
#6856: Improved exception handling within Rasa Open Source.

All exceptions that are somewhat expected (e.g. errors in file formats like
configurations or training data) will share a common base class
RasaException.

::warning Backwards Incompatibility
Base class for the exception raised when an action can not be found has been changed
from a NameError to a ValueError.
::

Some other exceptions have also slightly changed:
- raise YamlSyntaxException instead of YAMLError (from ruamel) when
  failing to load a yaml file with information about the line where loading failed
- introduced MissingDependencyException as an exception raised if packages
  need to be installed
#6900: Debug logs from matplotlib libraries are now hidden by default and are configurable with the LOG_LEVEL_LIBRARIES environment variable.
#6943: Update KafkaEventBroker to support SASL_SSL and PLAINTEXT protocols.

Bugfixes

#3597: Fixed issue where temporary model directories were not removed after pulling from a model server.

If the model pulled from the server was invalid, this could lead to large amounts of local storage usage.
#5038: Fixed a bug in the CountVectorsFeaturizer which resulted in the very first
message after loading a model to be processed incorrectly due to the vocabulary
not being loaded yet.
#5135: Fixed Rasa shell skipping button messages if buttons are attached to
a message previous to the latest.
#5385: Stack level for FutureWarning updated to level 2.
#5453: If custom utter message contains no value or integer value, then it fails
returning custom utter message. Fixed by converting the template to type string.
#5617: Don't create TensorBoard log files during prediction.
#5638: Fixed DIET breaking with empty spaCy model.
#5737: Pinned the library version for the Azure
Cloud Storage to 2.1.0 since the
persistor is currently not compatible with later versions of the azure-storage-blob
library.
#5755: Remove clean_up_entities from extractors that extract pre-defined entities.
Just keep the clean up method for entity extractors that extract custom entities.
#5792: Fixed issue where the DucklingHTTPExtractor component would
not work if its url contained a trailing slash.
#5808: Changed to variable CERT_URI in hangouts.py to a string type
#5850: Slots will be correctly interpolated for button responses.

Previously this resulted in no interpolation due to a bug.
#5905: Remove option token_pattern from CountVectorsFeaturizer.
Instead all tokenizers now have the option token_pattern.
If a regular expression is set, the tokenizer will apply the token pattern.
#5921: Allow user to retry failed file exports in interactive training.
#5964: Fixed a bug when custom metadata passed with the utterance always restarted the session.
#5998: WhitespaceTokenizer does not remove vowel signs in Hindi anymore.
#6042: Convert entity values coming from DucklingHTTPExtractor to string
during evaluation to avoid mismatches due to different types.
#6053: Update FeatureSignature to store just the feature dimension instead of the
complete shape. This change fixes the usage of the option share_hidden_layers
in the DIETClassifier.
#6087: Unescape the \n, \t, \r, \f, \b tokens on reading nlu data from markdown files.

On converting json files into markdown, the tokens mentioned above are espaced. These tokens need to be unescaped on loading the data from markdown to ensure that the data is treated in the same way.
#6120: Fix the way training data is generated in rasa test nlu when using the -P flag.
Each percentage of the training dataset used to be formed as a part of the last
sampled training dataset and not as a sample from the original training dataset.
#6143: Prevent WhitespaceTokenizer from outputting empty list of tokens.
#6198: Add EntityExtractor as a required component for EntitySynonymMapper in a pipeline.
#6222: Better handling of input sequences longer than the maximum sequence length that the HFTransformersNLP models can handle.

During training, messages with longer sequence length should result in an error, whereas during inference they are
gracefully handled but a debug message is logged. Ideally, passing messages longer than the acceptable maximum sequence
lengths of each model should be avoided.
#6231: When using the DynamoTrackerStore, if there are more than 100 DynamoDB tables, the tracker could attempt to re-create an existing table if that table was not among the first 100 listed by the dynamo API.
#6282: Fixed a deprication warning that pops up due to changes in numpy
#6291: Update rasabaster to fix an issue with syntax highlighting on "Prototype an Assistant" page.

Update default stories and rules on "Prototype an Assistant" page.
#6419: Fixed a bug in the serialise method of the EvaluationStore class which resulted in a wrong end-to-end evaluation of the predicted entities.
#6535: Forms with slot mappings defined in domain.yml must now be a
dictionary (with form names as keys). The previous syntax where forms was simply a
list of form names is still supported.
#6577: Remove BILOU tag prefix from role and group labels when creating entities.
#6601: Fixed a bug in the featurization of the boolean slot type. Previously, to set a slot value to "true",
you had to set it to "1", which is in conflict with the documentation. In older versions true
(without quotes) was also possible, but now raised an error during yaml validation.
#6603: Fixed a bug in rasa interactive. Now it exports the stories and nlu training data as yml file.
#6711: Fixed slots not being featurized before first user utterance.

Fixed AugmentedMemoizationPolicy to forget the first action on the first going back
#6741: Fixed the remote URL of ConveRT model as it was recently updated by its authors.
#6755: Treat the length of OOV token as 1 to fix token align issue when OOV occurred.
#6757: Fixed the bug when entity was extracted even
if it had a role or group but roles or groups were not expected.
#6803: Fixed the bug that caused supported_language_list of Component to not work correctly.

To avoid confusion, only one of supported_language_list and not_supported_language_list can be set to not None now
#6897: Fixed issue where responses including text: "" and no custom key would incorrectly fail domain validation.
#6898: Fixed issue where extra keys other than title and payload inside of buttons made a response fail domain validation.
#6919: Do not filter training data in model.py but on component side.
#6929: Check if a model was provided when executing rasa test core.
If not, print a useful error message and stop.
#6805: Transfer only response templates for retrieval intents from domain to NLU Training Data.

This avoids retraining the NLU model if one of the non retrieval intent response templates are edited.

Improved Documentation

#4441: Added documentation on ambiguity_threshold parameter in Fallback Actions page.
#4605: Remove outdated whitespace tokenizer warning in Testing Your Assistant documentation.
#5640: Updated Facebook Messenger channel docs with supported attachment information
#5675: Update rasa shell documentation to explain how to recreate external
channel session behavior.
#5811: Event brokers documentation should say url instead of host.
#5952: Update rasa init documentation to include tests/conversation_tests.md
in the resulting directory tree.
#6819: Update "Validating Form Input" section to include details about
how FormValidationAction class makes it easier to validate form slots in custom actions and how to use it.
#6823: Update the examples in the API docs to use YAML instead of Markdown

Miscellaneous internal changes

#5784, #5788, #6199, #6403, #6735

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

2.0.0

Deprecations and Removals

Features

Improvements

Bugfixes

Improved Documentation

Miscellaneous internal changes