Events Add an event Speakers Talks Collections Sign out
 
Duration 22:49
16+
Play
Video

Rasa Open Source - What’s next? | Rasa Summit 2021

Tom Bocklisch
Director of Engineering at Rasa
  • Video
  • Table of contents
  • Video
Rasa Summit 2021
February 12, 2021, Online, USA
Rasa Summit 2021
Request Q&A
Video
Rasa Open Source - What’s next? | Rasa Summit 2021
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
310
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Tom Bocklisch
Director of Engineering at Rasa

My goal is to empower people. I've Co-Founded a software agency where we are using cutting edge technologies to realize almost impossible ideas. Driven by the fascination of technology and data science, I am improving my knowledge constantly.

View the profile

About the talk

Presented by Rasa Director of Engineering Tom Boklisch at the 2021 Rasa Summit. Tom shared what's new and what's next for Rasa Open Source.

#conversationalai #aichatbots #opensource

- Learn more about Rasa: [https://rasa.com​​](https://www.youtube.com/redirect?even...​)

- Rasa documentation: [http://rasa.com/docs​​](https://www.youtube.com/redirect?even...​)

- Join the Rasa Community: [https://forum.rasa.com​​](https://www.youtube.com/redirect?even...​)

- Twitter: [https://twitter.com/Rasa_HQ​​](https://www.youtube.com/redirect?even...​)

- Facebook: [https://www.facebook.com/RasaHQ​​](https://www.youtube.com/redirect?even...​)

- Linkedin: [https://www.linkedin.com/company/rasa​​](https://www.youtube.com/redirect?even...​)

Share

All right. Welcome everyone. I'm Tom from Rhonda and but come to the Thursday or Friday. And I'm going to walk you through what we've been working on. What we've been working on this Raza. It's all right there. An open-source. What's next? We're going to talk a little bit about what we've been working on for the past couple months since the last major release. And it would kind of take, we're going to go on that journey and going to take a look at what's next, what we're currently working

on the 80s that we have. And, yeah, we really wanted your feedback on that we want during for. It's like, if you, if they decide that you want to share this, if you have other ideas, if you are researching some of these ideas and have Brayden ready to reach out to outside of knowing the booze, or an email or in enough for Rooms To Go to hear your thoughts on this. So we've released kind of our past major release to find out in October and you're currently sitting on 2.3 on which I think we've

released a couple hours ago. So we fresh off the press and I forgot we're going to take a look at the upcoming releases and you will see a difference between a parallel and we might release them in a different order. They are mentioned here, so it's really just more visualization to make them all fit on the stage. Okeechobee. Take a look at what's next. Let's take a step back. And I want to really make sure that you know, about all the features that we've shipped in the past couple months, I'm tired of O 2.0 release

the ship, the root policy, which is kind of a big step for us to kind of unify how we treat deterministic rules and kind of combined them with internal policies. This was a huge relief and even mention all of them. But I want to take a special look at the new trainee did it because it's super important for some of the fetus ever record. You watching I'm so you can you can see you have a conversation. I'm looking for a restaurant, have a Chinese food, sure. Okay. Looks like in our training day

you can see it's in llamo, it's kind of traditional he almost fired with sintex highlighting. You can see down at the bottom, the stories. And this particular story of Britain in that new format, So why did we do that? Why did you change training data format? And I mean, it was a huge step for us. First of all, to come up with a new format. Also making sure there's a good migration paths and we know that it's a huge, huge hassle for everyone to migrate their training day. So why did we come up with a new training

data format in the first place? So young tender traded to has a lot of benefits so we can get good editor, support. People know you almost have to find Yamaha and it is just a unified extensible format. So the training data format for a new stories is a lot more similar now and which analysis to do some cool things that were truly working on. And it's also a lot more extensive. Also, if we think about new features that we need to keep track of and need to annotate in our training data, we

can do that with the yellow format and we just hit limits with the old train that I forgot. And that was kind of the starting point for us to think about this new format. It's a really kind of making its future proof and being able to add more more data, Ashley progress than just keeping track of more information that we need to just throw the greatest. Another great feature is it composable? I'm so you can spit up your demesne, you can stick up your butt and your sister into multiple pieces and then combine them been training a motor and hopefully you've seen Adam Stark and its

keynote at the beginning of the first state and Tweedy sing. This is the future being able to split up and about and I combined that when building a model, it makes a huge. If you're writing a huge assistant that covers lots of a lot of crimes, lots of different domains and it's way easier to maintain that overtime. If you're able to clearly separate pieces and it combined them back again to train your final model. So this is why the yellow train sets us up for the future. In Reading allows us to

where can I research topics? But also allow you to build better but you can make it Another thing that we should this new documentation, and hopefully a lot, a lot of people already saw that, but we try to aim for better structure were complete documentation and I want to highlight one feature that I think it's super cool and that I'm sure we're going to make more use of in the future as well as didn't know, in Saw Creek Side until we have that playground in there that allows you to work on training data in your browser, training model, right in there, and

then talk to that. Assisted in, in your product right now, I need to install anything you can easily demo bought, you can easily share something real quick, or co-worker and show them to send texts until we reach, you know, this new feature on the playground in the documentation And I'm sure you're going to expand on that and make that even better in the future. If you have seen a link at the bottom into the docks, In another release that we did. I think, right before Christmas, we should think of, mental training. And I'm highlighting that because it allows us

to do a couple more things but let's take a quick look at how it works for. You have your original data are you going to train them all? Oh and using a hundred epochs so kind of training as long as needed to kind of get that original model after that you can imitate new data hopefully using real conversation. So greedy do you use them in your conversations and its training just for Jenny ethics, right? So I'm not 30% of not kind of like a just a fraction of it and you knew find you tomorrow though just

raining a lot faster than the original And this will allow you to write Quaker. Allow you to train a model act, annotating a couple of examples, Entre, nous model, and then use that to to each other industry. Using that train times. Huge benefit in the United Nation, cycle training. So we'll see what we're going to do with that where you can modify it at all. And ideally you could really see that new model and action. Elegant composability, right? If you think about huge demands about Bots, that spend so many times that you can't even count them in a single-file.

You need to screw that up and I miss you spit it out. Thank you change a small piece in one of the tires, if it require you to retrain the whole motor and making our joining an incremental training allows us to reach. You start burden on the training time that you would encounter there. So you could change something in a small part of your bought and just train and finds you not model lost quicker than you would need to do otherwise, right? If you would want to train a whole new model. Chicken right, super helpful for our for large pots. Really want to come Thursday

pieces together. Are we testing right now? So that's something that we should actually not tired of release the kind of outside, it's the kind of action and to ship is confidence. We want to allow you to automatically test an assistant to set up. This is so it should always test your assistant to make sure there's no regressions. You want to go out and have this allows you to be sure that you don't have any progression. So that's kind of standard practice. Until he saw that looks like. So you could create a new training, a new test story in

Raza, x copy that. I would copy that into your story sets and and you texted. And then on this case for guitar actions, we would train them all to run through all the tests or he's on a train across radiation model on Ellen. Post these results on a pull request, in this case, change the model of you. You've trained before and you will see if kind of your new model worse, if it's performing better and this allows you to see if there's a regression, are we shouldn't bench this Boulder Crest Road back? And then check why the change is resulting in

this? Savior. And we shipped these are kids get erections and you can check them out in the GitHub Marketplace and be sure there's going to be more of these for a difference because they kind of template and we can see if you're running on guitar, music, guitar actions. Do check them out and let us know what you think about them. Call Wright's last speech. I want to highlight is that we already ship sin, two point three. So I can set a couple hours ago is machine during inspection

and that's an interesting feature because it's not necessarily a surface level feature on, but it's something that we did under the urge to allow us to inspect models. And so what you can see, here is the rides and lit project. If you haven't checked out its streamlet applications that allow you to visualize certain parts of your models for training data and really allows you to get a dig deeper into your mother's and enjoy your day. I'm in what you can see here on the right hand side is it said, Police Inspector. And it shows you a lot more detail than

you would usually get by just looking at the Acura policy. I'm in this is just an example, but it kind of showcases that we can use this utilization. We didn't rise open source to allow you to customize the most easier, right? So you can spend that out, check out of what's going on under the hood with us and other people to do better research. As you can see, how does inputs change. For example of the potential dates, where do we see issues in the space and also help to be back in the future. I'm

so kind of being able to see you. Why something say it's a white Airship patterns of failure inside. All right. So that was a lot of talking about the toss time, but I really want some to spend some time talking about the future and talking about the features that we're currently working on. So what's next? And the first thing I want to highlight is identifying and successful conversations. It's Gail. It's been a problem that we've been looking into for quite some time and it gets more urgent as checkpoints in

the system's get more popular, write more users who use your assistant, but we still want to enable you to really look at actual conversations and learn from them. End. We want to enable you to understand. Where do you just drop off? Why do they drop off and what can you do to fix that? And on the one hand, but we want to do is enabled on in Raza X using more powerful practice inches is a free sample, is to review two conversations at fullback, actual Chico's. As a

fullback, is a good signal that something went wrong or if he takes search and allowing you to better, explore the whole set of conversations that you've collected from north of fuses. But at the same time, we also going to make it kind of future lies from the open-source to allow us to and if they were inturned to machine learning models and detect who wear something went wrong to do. What kind of working on the policy that recruited calling intend to Ted. And that will allow us to detect where an unexpected intent was expressed at which point in the conversation

that happened. So we'll be able to kind of see where we think something went wrong. And even if no fallback actual is triggered, right? Even though it is, kind of confidence is a good, this will help us to dig a little deeper and get more inside and switch positions, didn't go the right way. At the same time, also looking into continuous motivation and that's an idea that we think is super important to support on CD. So we talked in the Dubai testing earlier, but we also want to make sure that we're able to compare different brands of the bus. I'm despite the fact

that they might be working on different conversations and Steve went to Flight, 25 years of progress. And for example, we, we want to empower you to evaluate the impact of new features that you're heading to your boss newly, trained models, or maybe you're changing the components in your pipeline every month to be able to quantify the impact that has another example of Sandy testing, right? Let's say, you spend up to different parts, and we really want to be able to quantify, which one of these things better, even though they're kind of handing different

conversation, And I think that was a great step forward and being able to betray Don an assistant instead of making them better over time. A really big project is Breaking Free from intense and it's one that we've been working on and we think that's going to take a little bit more time to get that introduction at the same time. You think it's super important we thinking tends aren't the future but they still very necessary right now and there's going to be a time that we need to bridge where we want to use intense. And at the same time we want to move to that if assistance in the

right direction but never come if I run steaks designing and implementing hypethetic conversations training and synthetic data, engineering to put everything into a bucket and kind of labeling. Everything, that's what you're seeing you in. Czech Show to get to level 3 of Beyond, we have to reread our conversation, AI works and he's a conversation that you would usually expect. So I'm looking for a restaurant, I'll buy Chinese fruit sure. That that is conversation that you usually observe,

but what use is also do is I'm looking for a restaurant. How about Chinese food I had that yesterday? How about a change? Would you say that? But I know that, but it is tricky to do just intense. And so that's why we're exploring assistance that you can combine in French. Resistance that don't use and tents at all. And actually shoot something already and it's called experimental attention in training. But as you can see, even though we ship that as an experimental feature, we still need to do a bit more work to really get over the finish line and

make sure that this can be used on a larger scale. So improved versions of rava and you would send you a text to interview. You would get out any chance you could put that into dialogue manager and give it to. You can actually do both of these at the same time. So you would get intense, but he would also be able to send, a text directly due to the dialog manager and have that text Insurance the next election prediction. And Addison mobile. We want to explore and we sure that's a good direction but not necessarily

sure it's the final solution to that. You get to level four or five and so on. But we know that it's kind of an interesting direction that will enable our future use cases and Brittany to make this part of every assistant and it's just a heavy lifting to do. So we need to refactor quite a lot of components and parts of a traditionally separated things in the NAU score to go to find a messages and how do you want to be able to, to have this kind of shortcuts to post test

text? I like the unit manager, Go through all this refectory is that we want to do a video on a nav. To do more research, I guess. I do want to explore more of the rest of you search ideas that we have to really be sure that we're moving in the right direction and then if we need be able to kind of handle to use cases that we've just seen on the electric slide. Alright, last but not least, controllable energy and what do we want to achieve here? So we actually want to add Jeff the response to the context security that is possible. But you would need to write

custom Nordic for that and every customer that you have. So you could take a look at the history of the conversation and then programmatically select one temporary go together but it doesn't really still. So we want to change that. Add an addition to Dad over probably a bit of a longer-term. We also want to direct you learn from examples but in Dallas important part guarantee transparent control for conversational designers and developers. Randomly generates responses to

allow convert conversation of designers and developers to directions to the assistant and an important in the right direction. How does this? Looks like. So we need some, for example, is mirroring uses good choices. So, if you think about shoes, you can call and sneakers on us and kicks and a really cool. If your sister would be able to pick up the phrasing, the user uses, this should make sure that kind of the assistants more appealing for the body, and the user has a better experience, right? They feel understood because you kind of using the same language and

they are using Another example is interesting responses based on triviz responses. So if you think about a situation where you're unable to handle user request, the first time you would respond, I'm sorry. I don't understand. The second time you would say I'm sorry I didn't quite get what you mean. Would you like to speak to customer service at Star Trek to hand you over to customer service agent? This is already possible referral that you can do that. You can store that stayed and do that kind of programmatically in your code by

tvt. Want to make that easier and find a general pattern, there to allow you to use that and all your assistance and the different use cases. All right, so that's be over you. That's kind of what we still working on and really brief overview of kind of what's next. What's on our radar on. These are only some of the ideas, I couldn't cover all the things that we're working on. I think we still have 400 open issues on target about open source. This is really just a snapshot and I feel really happy to hear your thoughts

right to come and do and then you can choose any songs by to try to link a lot of source material right down, the slide show. Where I put the, I took that information from the exposed pipe issue. Sophia free to explore that more and come in and let us know what you think. All right, that thinks that I just said, lots of ideas. And if you're excited, also feel free to explore. Culligan.

Cackle comments for the website

Buy this talk

Access to the talk “Rasa Open Source - What’s next? | Rasa Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Rasa Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Artificial Intelligence and Machine Learning”?

You might be interested in videos from this event

October 7 - 20, 2020
Online, Mountain View
19
5.41 K
google, googledevs, it, machinelearning, mlsummit, network, platform, tensorflow, tfx

Similar talks

Hans Van Dam
Co-founder and CEO at Conversation Design Institute
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Nikhil Mane
Conversational AI Engineer at Autodesk
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Evgeniia Razumovskaia
Machine Learning Research Intern at Rasa
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “Rasa Open Source - What’s next? | Rasa Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
742 conferences
30258 speakers
11303 hours of content