Change Log¶

All notable changes to this project will be documented in this file. This project adheres to Semantic Versioning starting with version 0.2.0.

[0.13.0] - 2019-01-23¶

Added¶

A support for session persistence mechanism in the SocketIOInput compatible with the example SocketIO WebChat + short explanation on how session persistence should be implemented in a frontend
TwoStageFallbackPolicy which asks the user for their affirmation if the NLU confidence is low for an intent, for rephrasing the intent if they deny the suggested intent, and does finally an ultimate fallback if it does not get the intent right
Additional checks in PolicyEnsemble to ensure that custom Policy classes’ load function returns the correct type
Travis script now clones and tests the Rasa stack starter pack
Entries for tensorflow and sklearn versions to the policy metadata
SlackInput wont ignore app_mention event anymore. Will handle messages containing @mentions to bots and will respond to these (as long as the event itself is enabled in the application hosting the bot)
Added sanitization mechanism for SlackInput that (in its current shape and form) strips bot’s self mentions from messages posted using the said @mentions.
Added random seed option for KerasPolicy and EmbeddingPolicy to allow for reproducible training results
InvalidPolicyConfig error if policy in policy configuration could not be loaded, or if policies key is empty or not provided
Added a unique identifier to UserMessage and the UserUttered event.

Removed¶

support for deprecated intents/entities format

Changed¶

replaced pytest-pep8 with pytest-pycodestyle
switch from PyInquirer to questionary for the display of commandline interface (to avoid prompt toolkit 2 version issues)
if NLU classification returned None in interactive training, directly ask a user for a correct intent
trigger fallback on low nlu confidence only if previous action is action_listen
updated docs for interactive learning to inform users of the --core flag
Change memoization policies confidence score to 1.1 to override ML policies

Fixed¶

fix error during interactive learning which was caused by actions which dispatched messages using dispatcher.utter_custom_message
re-added missing python-engineio dependency
fixed not working examples in examples/
strip newlins from messages so you don’t have something like “n/restartn”
properly reload domain when using /model endpoint to upload new model
updated documentation for custom channels to use the credentials.yml

[0.12.3] - 2018-12-03¶

Added¶

added scipy dependency (previously pulled in through keras)
added element representation for command-line output

Changed¶

improved button representation for custom buttons in command-line

Changed¶

randomized initial sender_id during interactive training to avoid loading previous sessions from persistent tracker stores

Removed¶

removed keras dependency, since keras_policy uses tf.keras

[0.12.2] - 2018-11-20¶

Fixed¶

argument handling on evaluate script
added basic sanitization during visualization

[0.12.1] - 2018-11-11¶

Fixed¶

fixed interactive learning to properly submit executed actions to the action server
allow the specification of the policy configuration while using the visualisation script
use default configuration if no policy configuration is passed
fixed html delivery from interactive server script (package compatible)
SlackBot when created in SlackInputChannel inherits the slack_channel property, allowing Slack bots to post to any channel instead of only back to the user
fix writing of new domain file from interactive learning
fix reading of state featurizers from yaml
fix reading of batch_size parameter in keras policy

[0.12.0] - 2018-11-11¶

Warning

This is major new version with a lot of changes under the hood as well as on the API level. Please take a careful look at the Migration Guide guide before updating. You need to retrain your models.

Added¶

new connector for the Cisco Webex Teams chat
openapi documentation of server API
NLU data learned through interactive learning will now be stored in a separate markdown-format file (any previous NLU data is merged)
Command line interface for interactive learning now displays policy confidence alongside the action name
added action prediction confidence & policy to ActionExecuted event
the Core policy configuration can now be set in a config.yaml file. This makes training custom policies possible.
both the date and the time at which a model was trained are now included in the policy’s metadata when it is persisted
show visualization of conversation while doing interactive learning
option for end-to-end evaluation of Rasa Core and NLU examples in evaluate.py script
/conversations/{sender_id}/story endpoint for returning the end-to-end story describing a conversation
docker-compose file to start a rasa core server together with nlu, an action server, and duckling
http server (rasa_core.run --enable-api) evaluation endpoint
ability to add tracker_store using endpoints.yml
ability load custom tracker store modules using the endpoints.yml
ability to add an event broker using an endpoint configuration file
raise an exception when server.py is used instead of rasa_core.run --enable-api
add documentation on how to configure endpoints within a configuration file
auth_source parameter in MongoTrackerStore defining the database to authenticate against
missing instructions on setting up the facebook connector
environment variables specified with ${env_variable} in a yaml configuration file are now replaced with the value of the environment variable
detailed documentation on how to deploy Rasa with Docker
make wait_time_between_pulls configurable through endpoint configuration
add FormPolicy to handle form action prediction
add ActionExecutionRejection exception and ActionExecutionRejected event
add default action ActionDeactivateForm()
add formbot example
add ability to turn off auto slot filling with entity for each slot in domain.yml
add InvalidDomain exception
add active_form_... to state dictionary
add active_form and latest_action_name properties to DialogueStateTracker
add Form and FormValidation events
add REQUESTED_SLOT constant
add ability to read action_listen from stories
added train/eval scripts to compare policies

Changed¶

improved response format for /predict endpoint
all error messages from the server are now in json format
agent.log_message now returns a tracker instead of the trackers state
the core container does not load the nlu model by default anymore. Instead it can be connected to a nlu server.
stories are now visualized as .html page instead of an image
move and deduplicate restaurantbot nlu data from franken_data.json to nlu_data.md
forms were completely reworked, see changelog in rasa_core_sdk
state featurization if some form is active changed
Domain raises InvalidDomain exception
interactive learning is now started with rasa_core.train interactive
passing a policy config file to train a model is now required
flags for output of evaluate script have been merged to one flag --output where you provide a folder where any output from the script should be stored

Removed¶

removed graphviz dependency
policy config related flags in training script (see migration guide)

Fixed¶

fixed an issue with boolean slots where False and None had the same value (breaking model compatibility with models that use a boolean slot)
use utf8 everywhere when handling file IO
argument --connector on run script accepts custom channel module names
properly handle non ascii categorical slot values, e.g. 大于100亿元
fixed HTTP server attempting to authenticate based on incorrect path to the correct JWT data field
all sender ids from channels are now handled as str. Sender ids from old messages with an int id are converted to str.
legacy pep8 errors

[0.11.12] - 2018-10-11¶

Changed¶

Remove livechat widget from docs

[0.11.11] - 2018-10-05¶

Fixed¶

Add missing name() to facebook Messenger class

[0.11.10] - 2018-10-05¶

Fixed¶

backport fix to JWT schema

[0.11.9] - 2018-10-04¶

Changed¶

pin tensorflow 1.10.0

[0.11.8] - 2018-09-28¶

Fixed¶

cancel reminders if there has been a restarted event after the reminder

Changed¶

JWT authentication now checks user roles. The admin role may access all endpoints. For endpoints which contain a sender_id parameter, users with the user role may only call endpoints where the sender_id matches the user’s username.

[0.11.7] - 2018-09-26¶

Added¶

custom message method in rocketchat channel

Fixed¶

don’t fail if rasa and rest input channels are used together
wrong paramter name in rocketchat channel methods
Software 2.0 link on interactive learning documentation page went to Tesla’s homepage, now it links to Karpathy blogpost

[0.11.6] - 2018-09-20¶

Added¶

UserMessage and UserUttered classes have a new attribute input_channel that stores the name of the InputChannel through which the message was received

[0.11.5] - 2018-09-20¶

Fixed¶

numpy version incompatibility between rasa core and tensorflow

[0.11.4] - 2018-09-19¶

Added¶

a flag --fail_on_prediction_errors to the evaluate.py script - if used when running the evaluation, the script will fail with a non 0 exit code if there is at least one prediction error. This can be used on CIs to validate models against test stories.
JWT support: parameters to allow clients to authenticate requests to the rasa_core.server using JWT’s in addition to normal token based auth
added socket.io input / output channel
UserMessage and UserUttered classes have a new attribute input_channel that stores the name of the InputChannel through which the message was received

Changed¶

dump failed stories after evaluation in the normal story format instead of as a text file
do not run actions during evaluation. instead, action are only predicted and validated against the gold story.
improved the online learning experience on the CLI
made finetuning during online learning optional (use --finetune if you want to enable it)

Removed¶

package pytest-services since it wasn’t necessary

Fixed¶

fixed an issue with the followup (there was a name confusion, sometimes the followup action would be set to the non existent follow_up_action attribute instead of followup_action)

[0.11.3] - 2018-09-04¶

Added¶

callback output channel, receives messages and uses a REST endpoint to respond with messages

Changed¶

channel input creation moved to the channel, every channel can now customize how it gets created from the credentials file

[0.11.2] - 2018-09-04¶

Changed¶

improved documentation for events (e.g. including json serialisation)

Removed¶

outdated documentation for removed endpoints in the server (/parse & /continue)

Fixed¶

read in fallback command line args

[0.11.1] - 2018-08-30¶

Fixed¶

increased minimal compatible model version to 0.11.0

[0.11.0] - 2018-08-30¶

Warning

This is major new version with a lot of changes under the hood as well as on the API level. Please take a careful look at the Migration Guide guide before updating. You need to retrain your models.

Added¶

added microsoft botframework input and output channels
added rocket chat input and output channels
script parameter --quiet to set the log level to WARNING
information about the python version a model has been trained with to the model metadata
more emoji support for PY2
intent confidence support in RegexInterpreter
added paramter to train script to pull training data from an url instead of a stories file
added new policy: Embedding Policy implemented in tensorflow

Changed¶

default log level for all scripts has been changed from WARNING to INFO.
format of the credentials file to allow specifying the credentials for multiple channels
webhook URLs for the input channels have changed and need to be reset
deprecated using rasa_core.server as a script - use rasa_core.run --enable_api instead
collecting output channel will no properly collect events for images, buttons, and attachments

Removed¶

removed the deprecated TopicSet event
removed tracker.follow_up_action - use the FollowupAction event instead
removed action_factory: remote from domain file - the domain is always run over http
removed OnlineLearningPolicy - use the training.online script instead

Fixed¶

lots of type annotations
some invalid documentation references
changed all logger.warn to logger.warning

[0.10.4] - 2018-08-08¶

Added¶

more emoji support for PY2
intent confidence support in RegexInterpreter

[0.10.3] - 2018-08-03¶

Changed¶

updated to Rasa NLU 0.13
improved documentation quickstart

Fixed¶

server request argument handling on python 3
creation of training data story graph - removes more nodes and speeds up the training

[0.10.2] - 2018-07-24¶

Added¶

new RasaChatInput channel
option to ignore entities for certain intents

Fixed¶

loading of NLU model

[0.10.1] - 2018-07-18¶

Changed¶

documentation changes

[0.10.0] - 2018-07-17¶

Warning

This is a major new release with backward incompatible changes. Old trained models can not be read with the new version - you need to retrain your model. View the Migration Guide for details.

Added¶

allow bot responses to be managed externally (instead of putting them into the domain.yml)
options to prevent slack from making re-deliver message upon meeting failure condition. the default is to ignore http_timeout.
added ability to create domain from yaml string and export a domain to a yaml string
added server endpoint to fetch domain as json or yaml
new default action ActionDefaultFallback
event streaming to a RabbitMQ message broker using Pika
docs section on event brokers
Agent() class supports a model_server EndpointConfig, which it regularly queries to fetch dialogue models
this can be used with rasa_core.server with the --endpoint option (the key for this the model server config is model)
docs on model fetching from a URL

Changed¶

changed the logic inside AugmentedMemoizationPolicy to recall actions only if they are the same in training stories
moved AugmentedMemoizationPolicy to memoization.py
wrapped initialization of BackgroundScheduler in try/except to allow running on jupyterhub / binderhub/ colaboratory
fixed order of events logged on a tracker: action executed is now always logged before bot utterances that action created

Removed¶

removed support for topics

[0.9.6] - 2018-06-18¶

Fixed¶

fixed fallback policy data generation

[0.9.5] - 2018-06-14¶

Fixed¶

handling of max history configuration in policies
fixed instantiation issues of fallback policy

[0.9.4] - 2018-06-07¶

Fixed¶

fixed evaluation script
fixed story file loading (previously some story files with checkpoints could create wrong training data)
improved speed of data loading

[0.9.3] - 2018-05-30¶

Fixed¶

added token auth to all endpoints of the core server

[0.9.2] - 2018-05-30¶

Fixed¶

fix handling of max_history parameter in AugmentedMemoizationPolicy

[0.9.1] - 2018-05-29¶

Fixed¶

persistence of training data collected during online learning if default file path is used
the agent() method used in some rasa_core.server endpoints is re-run at every new call of the ensure_loaded_agent decorator
fixed OR usage of intents

[0.9.0] - 2018-05-24¶

Warning

This is a major new release with backward incompatible changes. Old trained models can not be read with the new version - you need to retrain your model.

Added¶

supported loading training data from a folder - loads all stories from all files in that directory
parameter to specify NLU project when instantiating a RasaNLUInterpreter
simple /respond endpoint to get bot response to a user message
/conversations endpoint for listing sender ids of running conversations
added a Mattermost channel that allows Rasa Core to communicate via a Mattermost app
added a Twilio channel that allows Rasa Core to communicate via SMS
FallbackPolicy for executing a default message if NLU or core model confidence is low.
FormAction class to make it easier to collect multiple pieces of information with fewer stories.
Dockerfile for rasa_core.server with a dialogue and Rasa NLU model

Changed¶

moved server from klein to flask
updated dependency fbmessenger from 4.3.1 to 5.0.0
updated Rasa NLU to 0.12.x
updated all the dependencies to the latest versions

Fixed¶

List slot is now populated with a list
Slack connector: slack_channel kwarg is used to send messages either back to the user or to a static channel
properly log to a file when using the run script
documentation fix on stories

[0.8.6] - 2018-04-18¶

Fixed¶

pin rasa nlu version to 0.11.4 (0.12.x only works with master)

[0.8.5] - 2018-03-19¶

Fixed¶

updated google analytics docs survey code

[0.8.4] - 2018-03-14¶

Fixed¶

pin pykwalify<=1.6.0 as update to 1.6.1 breaks compatibility

[0.8.3] - 2018-02-28¶

Fixed¶

pin fbmessenger version to avoid major update

[0.8.2] - 2018-02-13¶

Added¶

script to reload a dumped trackers state and to continue the conversation at the end of the stored dialogue

Changed¶

minor updates to dependencies

Fixed¶

fixed datetime serialisation of reminder event

[0.8.1] - 2018-02-01¶

Fixed¶

removed deque to support python 3.5
Documentation improvements to tutorials
serialisation of date time value for ReminderScheduled event

[0.8.0] - 2018-01-30¶

This is a major version change. Make sure to take a look at the Migration Guide in the documentation for advice on how to update existing projects.

Added¶

--debug and --verbose flags to scripts (train.py, run.py, server.py) to set the log level
support for story cycles when using checkpoints
added a new machine learning policy SklearnPolicy that uses an sklearn classifier to predict actions (logistic regression by default)
warn if action emits events when using a model that it did never emit in any of the stories the model was trained on
support for event pushing and endpoints to retrieve the tracker state from the server
Timestamp to every event
added a Slack channel that allows Rasa Core to communicate via a Slack app
added a Telegram channel that allows Rasa Core to communicate via a Telegram bot

Changed¶

rewrite of the whole FB connector: replaced pymessenger library with fbmessenger
story file utterance format changed from * _intent_greet[name=Rasa] to * intent_greet{"name": "Rasa"} (old format is still supported but deprecated)
persist action names in domain during model persistence
improved travis build speed by not using miniconda
don’t fail with an exception but with a helpful error message if an utterance template contains a variable that can not be filled
domain doesn’t fail on unknown actions but emits a warning instead. this is to support reading logs from older conversation if one recently removed an action from the domain

Fixed¶

proper evaluation of stories with checkpoints
proper visualisation of stories with checkpoints
fixed float slot min max value handling
fixed non integer feature decoding, e.g. used for memoization policy
properly log to specified file when starting Rasa Core server
properly calculate offset of last reset event after loading tracker from tracker store
UserUtteranceReverted action incorrectly triggered actions to be replayed

[0.7.9] - 2017-11-29¶

Fixed¶

visualisation using Networkx version 2.x
add output about line of failing intent when parsing story files

[0.7.8] - 2017-11-27¶

Fixed¶

Pypi readme rendering

[0.7.7] - 2017-11-24¶

Added¶

log bot utterances to tracker

Fixed¶

documentation improvements in README
renamed interpreter argument to rasa core server

[0.7.6] - 2017-11-15¶

Fixed¶

moodbot example train command in docs

[0.7.5] - 2017-11-14¶

Changed¶

“sender_id” (and “DEFAULT_SENDER_ID”) keyword consistency issue #56

Fixed¶

improved moodbot example - more nlu examples as well as better fitting of dialogue model

[0.7.4] - 2017-11-09¶

Changed¶

added method to tracker to retrieve the latest entities #68

[0.7.3] - 2017-10-31¶

Added¶

parameter to specify font size when rendering story visualization

Fixed¶

fixed documentation of story visualization

[0.7.2] - 2017-10-30¶

Added¶

added facebook bot example
added support for conditional checkpoints. a checkpoint can be restricted to only allow one to use it if certain slots are set. see docs for details
utterance templates in domain yaml support buttons and images
validate domain yaml and raise exception on invalid file
run script to load models and handle messages from an input channel

Changed¶

small dropout in standard keras model to decrease reliance on exact intents
a LOT of documentation improvements

Fixed¶

fixed http error if action listen is not confirmed. #42

[0.7.1] - 2017-10-06¶

Fixed¶

issues with restart events. They created wrong a messed up history leading to wrong predictions

[0.7.0] - 2017-10-04¶

Added¶

support for Rasa Core usage as a server with remote action execution

Changed¶

switched to max code line length 80
removed action id - use action.name() instead. if an action implementation overrides the name, it should include the action_ prefix (as it is not automatically added anymore)
renamed rasa_dm.util to rasa_dm.utils
renamed the whole package to rasa_core (so rasa_dm is gone!)
renamed Reminder attribute id to name
a lot of documentation improvements. docs are now at https://rasa.com/docs/core
use hashing when writing memorized turns into persistence - requires retraining of all models that are trained with a version prior to this
changed agent.handle_message(...) interface for easier usage

[0.6.0] - 2017-08-27¶

Added¶

support for multiple policies (e.g. one memoization and a Keras policy at the same time)
loading domains from yaml files instead of defining them with python code
added an api layer (called Agent) for you to use for 95% of the things you want to do (training, persistence, loading models)
support for reminders

Changed¶

large refactoring of code base

[0.5.0] - 2017-06-18¶

Added¶

ScoringPolicy added to policy implementations (less strict than standard default policy)
RasaNLUInterpreter to run a nlu instance within dm (instead of using the http interface)
more tests

Changed¶

UserUtterance now holds the complete parse data from nlu (e.g. to access attributes other than entities or intent)
Turn has a reference to a UserUtterance instead of directly storing intent & entities (allows access to other data)
Simplified interface of output channels
order of actions in the DefaultPolicy in possible_actions (ActionListen now always has index 0)

Fixed¶

RedisTrackerStore checks if tracker is stored before accessing it (otherwise a None access exception is thrown)
RegexInterpreter checks if the regex actually matches the message instead of assuming it always does
str implementation for all events
Controller can be started without an input channel (e.g. messages need to be fed into the queue manually)

[0.2.0] - 2017-05-18¶

First released version.

Changed¶

Change payloads from “text” to “message” in files: server.yml, docs/connectors.rst, rasa_core/server.py, rasa_core/training/interactive.py, tests/test_interactive.py