Rasa Open Source - What’s next?
Rasa Summit 2021
Tom Bocklisch, Director of Engineering @ Rasa
February 2021
Rasa Open Source
New Training Data Format
Rule Policy
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
YAML Training Data Format
Rasa Open Source 2.0
https://rasa.com/docs/rasa/training-data-format
YAML Training Data Format
Rasa Open Source 2.0
Why did we do this?
● Well-known standard
● Good editor support
● Unified extensible format
● Composable
money_transfer:
- I want to transfer money
- I want to make a transfer
Financial bot
loan_queries:
- I need a loan
- I want to take out a loan
https://rasa.com/docs/rasa/training-data-format
Rasa Open Source
New Training Data Format
Rule Policy
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
New Documentation
New Documentation including
a brand new Playground
Rasa Open Source 2.0
Why did we do this?
● Better structure
● Completeness
● No-install quick start
https://rasa.com/docs/rasa/playground
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
New Documentation
Incremental Training
Rasa Open Source 2.2
https://blog.rasa.com/rasa-new-incremental-training/
Why did we do this?
● Iterate quickly
● On-the-spot retraining
● Composability
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
New Documentation
Automated Testing
Automated Testing of Assistants
Rasa Open Source 2.2
Why did we do this?
● Ship with confidence
● Easy to set up
● Follow SE best practices
https://github.com/marketplace?query=rasahq
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
New Documentation
Automated Testing
Machine Learning Model Inspection
Rasa Open Source 2.3
Why did we do this?
● Customization
● Research
● Debuggability
https://github.com/rasahq/rasalit
Future of Rasa Open Source
So, what is next then?
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
New Documentation
Automated Testing
Where do users drop off? Why do users drop off? What can I do to fix that?
Identifying Unsuccessful Conversations at Scale
FEATURE PREVIEW
More powerful, proactive filters
(e.g. notifications to review new conversations where a
fallback action was triggered, free text search, etc.)
Incorporate emerging research (IntentTED)
(e.g. detect where an unexpected intent was expressed at a point
in the conversation)
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
Continuous Model
Evaluation
New Documentation
Automated Testing
Continuous Model Evaluation methodology for CDD
FEATURE PREVIEW
What do we want to achieve?
● The ability to compare versions of
a bot despite the fact that they
participated in different
conversations.
● Providing a quantifiable measure
of progress.
● Estimating and evaluating the
impact of a new
feature/model/model component.
Second Bot
Version
First Bot
Version
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
Continuous Model
Evaluation
Breaking free
from Intents
New Documentation
Automated Testing
Adaptive Assistants will never come if everyone
sticks with designing and implementing
hypothetical conversations, training on synthetic
data, etc.
To get to level 3 and beyond we have to reinvent how conversational AI works.
LEVEL 3 AND BEYOND
What you expect: What users also do:
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
Continuous Model
Evaluation
Breaking free
from Intents
Experimental
e2e Training
New Documentation
Automated Testing
Get end-to-end ready for prime time. Breaking free of intents is just the start.
FEATURE PREVIEW
All of these changes will make it easier to modify models or integrate research ideas.
We are aiming to make end-to-end
part of every assistant.
Still some heavy lifting ahead:
● Refactor Structure of
Dialogue Events
● Refactor Tracker Store
Persistence
● Refactor Agent / Processor /
Tracker Store
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
Continuous Model
Evaluation
Breaking free
from Intents
Experimental
e2e Training
Controllable NLG
New Documentation
Automated Testing
Controllable Natural Language Generation (NLG)
FEATURE PREVIEW
What do we want to achieve?
● Adapt responses to their context →
currently possible with custom logic in
an action → does not scale
● NLG model that learns from examples
but guarantees transparent control
for conversational designers and
developers (longer term)
What does this look like?
1. Mirror users word choices
● If user uses “sneakers” assistant uses “sneakers”
● If user uses “runners” assistant uses “runners”
● If user uses “kicks” assistant uses “kicks”
2. Adjusting response based on previous responses
● First time: “I’m sorry. I don’t understand. Could you
try rephrasing?”
● Second time: “I’m very sorry. I don’t quite get what
you mean. Would you like to speak to customer
service?”
● Third time: “My apologies. Let’s hand you over to a
customer service agent, ...”
Rasa Open Source
New Training Data Format
Rule Policy
Incremental Training
Last Major Release
2.0
Current Version
2.3
Upcoming
Releases
Machine Learning
Model Inspection
Identifying Unsuccessful
Conversations at Scale
Continuous Model
Evaluation
Breaking free
from Intents
Experimental
e2e Training
Controllable NLG
New Documentation
Automated Testing
Tom Bocklisch
Director of Engineering
tom@rasa.com
@tombocklisch
Excited? Help us make
these ideas become reality!
rasa.com/jobs
👩‍💻 👨‍🚀 🚀

Rasa Open Source - What's next?

  • 1.
    Rasa Open Source- What’s next? Rasa Summit 2021 Tom Bocklisch, Director of Engineering @ Rasa February 2021
  • 2.
    Rasa Open Source NewTraining Data Format Rule Policy Last Major Release 2.0 Current Version 2.3 Upcoming Releases
  • 3.
    YAML Training DataFormat Rasa Open Source 2.0 https://rasa.com/docs/rasa/training-data-format
  • 4.
    YAML Training DataFormat Rasa Open Source 2.0 Why did we do this? ● Well-known standard ● Good editor support ● Unified extensible format ● Composable money_transfer: - I want to transfer money - I want to make a transfer Financial bot loan_queries: - I need a loan - I want to take out a loan https://rasa.com/docs/rasa/training-data-format
  • 5.
    Rasa Open Source NewTraining Data Format Rule Policy Last Major Release 2.0 Current Version 2.3 Upcoming Releases New Documentation
  • 6.
    New Documentation including abrand new Playground Rasa Open Source 2.0 Why did we do this? ● Better structure ● Completeness ● No-install quick start https://rasa.com/docs/rasa/playground
  • 7.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases New Documentation
  • 8.
    Incremental Training Rasa OpenSource 2.2 https://blog.rasa.com/rasa-new-incremental-training/ Why did we do this? ● Iterate quickly ● On-the-spot retraining ● Composability
  • 9.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases New Documentation Automated Testing
  • 10.
    Automated Testing ofAssistants Rasa Open Source 2.2 Why did we do this? ● Ship with confidence ● Easy to set up ● Follow SE best practices https://github.com/marketplace?query=rasahq
  • 11.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection New Documentation Automated Testing
  • 12.
    Machine Learning ModelInspection Rasa Open Source 2.3 Why did we do this? ● Customization ● Research ● Debuggability https://github.com/rasahq/rasalit
  • 13.
    Future of RasaOpen Source So, what is next then?
  • 14.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale New Documentation Automated Testing
  • 15.
    Where do usersdrop off? Why do users drop off? What can I do to fix that? Identifying Unsuccessful Conversations at Scale FEATURE PREVIEW More powerful, proactive filters (e.g. notifications to review new conversations where a fallback action was triggered, free text search, etc.) Incorporate emerging research (IntentTED) (e.g. detect where an unexpected intent was expressed at a point in the conversation)
  • 16.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale Continuous Model Evaluation New Documentation Automated Testing
  • 17.
    Continuous Model Evaluationmethodology for CDD FEATURE PREVIEW What do we want to achieve? ● The ability to compare versions of a bot despite the fact that they participated in different conversations. ● Providing a quantifiable measure of progress. ● Estimating and evaluating the impact of a new feature/model/model component. Second Bot Version First Bot Version
  • 18.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale Continuous Model Evaluation Breaking free from Intents New Documentation Automated Testing
  • 19.
    Adaptive Assistants willnever come if everyone sticks with designing and implementing hypothetical conversations, training on synthetic data, etc.
  • 20.
    To get tolevel 3 and beyond we have to reinvent how conversational AI works. LEVEL 3 AND BEYOND What you expect: What users also do:
  • 21.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale Continuous Model Evaluation Breaking free from Intents Experimental e2e Training New Documentation Automated Testing
  • 22.
    Get end-to-end readyfor prime time. Breaking free of intents is just the start. FEATURE PREVIEW All of these changes will make it easier to modify models or integrate research ideas. We are aiming to make end-to-end part of every assistant. Still some heavy lifting ahead: ● Refactor Structure of Dialogue Events ● Refactor Tracker Store Persistence ● Refactor Agent / Processor / Tracker Store
  • 23.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale Continuous Model Evaluation Breaking free from Intents Experimental e2e Training Controllable NLG New Documentation Automated Testing
  • 24.
    Controllable Natural LanguageGeneration (NLG) FEATURE PREVIEW What do we want to achieve? ● Adapt responses to their context → currently possible with custom logic in an action → does not scale ● NLG model that learns from examples but guarantees transparent control for conversational designers and developers (longer term) What does this look like? 1. Mirror users word choices ● If user uses “sneakers” assistant uses “sneakers” ● If user uses “runners” assistant uses “runners” ● If user uses “kicks” assistant uses “kicks” 2. Adjusting response based on previous responses ● First time: “I’m sorry. I don’t understand. Could you try rephrasing?” ● Second time: “I’m very sorry. I don’t quite get what you mean. Would you like to speak to customer service?” ● Third time: “My apologies. Let’s hand you over to a customer service agent, ...”
  • 25.
    Rasa Open Source NewTraining Data Format Rule Policy Incremental Training Last Major Release 2.0 Current Version 2.3 Upcoming Releases Machine Learning Model Inspection Identifying Unsuccessful Conversations at Scale Continuous Model Evaluation Breaking free from Intents Experimental e2e Training Controllable NLG New Documentation Automated Testing
  • 26.
    Tom Bocklisch Director ofEngineering tom@rasa.com @tombocklisch Excited? Help us make these ideas become reality! rasa.com/jobs 👩‍💻 👨‍🚀 🚀