back to episodes

The Friday Review - 2024-07-26

Discover the latest updates in ChatBotKit, including trigger integrations for autonomous AI agents and enhancements in model options with GPT-4o Mini. Explore new SDK features, playground tools, and community spaces like Slack. Learn how to build actionable bots with our tutorials and examples. Join us in pushing the boundaries of AI capabilities!

Transcript

We're going over some of the main changes. Hopefully, it will be educational, but also, will inspire you to try some of these features and get an idea of the, amount of work that we're putting in the ChatBotKit.

Two major changes this month, and the first one is called trigger integrations. I'd like you to think of it as the beginning of autonomous AI agents in the context of ChatBotKit. What trigger integrations allow you to do is to connect external systems into ChatBotKit and trigger bots based on the events received from those systems. In the context of a webhook, such as a webhook from Stripe or some other system, you will have that other system sending events to the bot, and the bot will be able to respond to those events.

Bot will be able to not only respond within the specific conversations, but also perform actions. Sky's the limit here the bot, the events, the skillset and abilities, we can create very complex autonomous systems.

At the moment, the trigger event feature is in beta.

The way you create triggers is relatively straightforward. All you have to do is just go to integrations and then choose the event trigger integration. You can specify the name of the description. You can choose the bot you can attach this integration to. If you don't choose a bot, you can just use the bot customization options and create what we call a shadow bot, but it's like a bot without an instance.

Let's create a bot just quickly for testing purposes. And let's create the integration. I'll demonstrate what this integration does at the very high levels. We're not going to go into some of the technical details and I promise that we're going to do a technical deep dive over the next week as well.

This is the most important bit. We have an event endpoint. This is where we're going to be sending the events and this is the secret. The secret is only used for authentication.

We grab the secret and we configure these two parameters into the system that is sending the events. We can provide some additional instructions in the text field. For example, if I want these events to be processed in a certain way, I can just describe how exactly.

Imagine if this is a Stripe and I want every time someone that subscribes send them an email. I can say something along the lines of " use the event to send a thank you email each subscriber", or something along those lines. That is an additional instruction.

We can also set up session duration. This is an option that is very similar to the ones that we already have with some of the other integrations. But what this provides is the window for which the conversation will be valid.

If you want to perform a complex operation where some of the actions that need be performed depend on other events received by the system, we can do so here. By keeping a longer context window, this means that the bot will be aware of all the events sent previously.

The next item on the list was the OpenAI GPT-4o Mini. Model is in beta state, but this is by far the cheapest model. Now, just because it's cheaper and because OpenAI advertises it as more powerful than GPT-3.5 Turbo, it doesn't necessarily mean that can be used everywhere. It really depends, and I advise to try the new model and see how it performs.

It will be hit and miss sometimes. Keep in mind that, the way this model scores is not representative, how it operate s. It might be the case that , GPT-4o Mini does not perform so well at specific tasks and you might find better performance using different models.

This is why a ChatBotKit have a pretty large collection of models that you can choose from. It's very important for you to try it before committing. It's very cost effective. It's very cheap. It's fast. But you need to also keep in mind that because it's new and it's so cost effective and it's so cheap, this means there'll be a lot more demand. We are anticipating that there will be service degradation when using this model. So this is a bit of a warning and this is why we still advertise it as a beta model.

We now allow you to use Groq as your own model. If you go to bots in the advanced options section, when you select the model, you can also select custom. We have different providers . We have OpenAI, Mistral, and now we have Groq.

We are still not advertising any Groq models at this stage, simply because there are certain limitations, in terms of how many requests you can make per minute. And unfortunately limits do not scale well with the amount of customers we have with the  demands of the models that we have at the moment. It will just won't work for us.

That's why we're not actively advertising any Groq models, but if you prefer to bring your own through your own account and your own keys, you can certainly do that by this configuration here.

The Node SDK was updated to version one 1.11.0. In this version, we are bringing some missing APIs, which were not exposed to the to the SDK in previous iterations: primarily secret management and contact management.

You can do that right now with the SDK as well as provide some additional methods for you to access better usage and analytics through the API. It is a very small update. You can find the SDK here. This is the official repository. We are planning to do some deep dives on the SDK in the coming days, which hopefully will give you a bit of an insight how to use it and how to set it up, especially in the context of production applications.

We've launched a Slack workspace. This is in addition to our Discord community and it's been frequently requested. Our Slack workspace is still in its infancy and it's not as well developed as the Discord one. That being said it's open. Some customers who might prefer to use Slack and we are open to support them. Feel free to join.

We also introduced a couple more playgrounds. Now, they're not directly associated with conversational AI, but they can certainly help in some cases. We've introduced three playgrounds: API, also JSONPath and JMESPath. This is for advanced use cases.

We have increased our fetch limits across all tiers. Now we have made it such as that you can perform as many fetch requests as there are messages available. This means that you can fully utilize the entire plan, especially in the combination with skillsets to create very advanced bots that are fetching data from The Internet, interacting with remote systems, APIs, so on and so forth. This is a very welcome update that's gonna open the doors to a lot more interesting solutions.

We have a tutorial that discusses the difference between the stateless and stateful interactions. So I definitely recommend this as a starting tutorial when you learn how to use the SDK, because that's going to help you understand how to use in the most appropriate way, depending on your specific use case. And also we have a pretty cool tutorial about how to make a Shopify AI bot. And it's just a step by step tutorial that demonstrates how to create a Shopify assistant. And the Shopify assistant will be able to find information about an order, get the status of the order and also help the customer choose a product. We also have a cool demo, which is a plug and play example, you can find here. This example has two types of actions, tracking order and product recommendation. It's pretty cool example and the actual tutorial goes over how to set it up.

And last but not least this month, we also have a new example in our example section, and this is the Move Recommendation ChatBot. It's a very simple example, but it just demonstrates how to use skillsets to interact with remote systems. And the main system is the movie database. The example contains two skillset abilities.

They allow us to query for trending movies of the week, make suggestions of movies that we would like to watch. The difference here using skillsets versus querying the model directly is that this information is 100% up to date because it's fetched fresh from the source and the source in this case is the movie database API.

The easiest way to test this example is to copy. You can just simply go to the example section, then create the bot and it will create all the relevant resources for you. Once the resources are created, you just need to configure only one piece and that is to set up a secret for the movie database API key. You need to have an account for that. Please grab an account from the movie database, enter the key, and you will be good to go.

This is also new. It's a small change where we are showing a solution diagram. You can see the different components such as the backstory and the model, the fact that produces conversation, the user can interact with this bot via widget, and the bot is connected to a skillset called Move Recommendations, and the skills have two abilities, Get Trending movies and Search M movies.

This part is designed for you to get a visual understanding of how different components are connected.

Thank you for listening and thank you for being here.

AI Widget