The Latest on Wagtail AI | What's New in Wagtail Recap

Speaker 1: I am going to talk quite briefly about white LAI. We had a webinar just on Wipped L AI. a couple of months ago, six weeks ago, and um it was really well attended. So some of you may have may have been there of that and uh in which case you'll you'll know what I'm talking about in the in the next summary, but then I'll show you a bit about what will work on so wagtail ai is uh is a third party app it's one that's we're developing actively at torchbox but uh with the help of the community it's not uh new default functionality in WhiteL. It's something that that you would opt into using by by installing it as a third-party app.

Speaker 1: And importantly, we uh we we're peggybacking on the LLM library, which uh has been open sourced by Simon Wilson, which uh is a sort of abstraction layer over other um large language models. So OpenAI's GPT API is of course one of the options and for many people it would be the default option but um it also supports Claude including the the new Claude 3 models and uh and open source models that that you might be hosting yourselves using Olama or or or on your on on other infrastructure. And I'll just show you quickly how Wagtail AI works at the moment. So

Speaker 1: At the moment the main way you interact with it is through a rich text field. So here is some text with some mistakes uh you might your spell your built-in browser spell checker might have noticed that one was wrong but probably not this because some is spelled correctly here and there's a grammar mistake so I can select that choose the wand and uh go for AI correction. So that's just making the call to the LM and with a with a prompt saying please fix this. There's also the option for generating content. So if I want to write a short post about Lombicer. But it there's two default prompts and I'll choose the second one and uh it's written out some text for me, most of which I probably wouldn't use, but could be a helpful starter.

Speaker 1: So just those two points themselves are a useful start. But what I want to highlight is how easy it is for you to write your own prompts. So if I go now to settings and then prompts , we delete the page and I'm going to add a new prompt. And I'm gonna ask it to rewrite the following text using very simple language, which could be understood by a five-year-old child. The description is explained by five. And then I here I choose the method, so whether I want to replace content or append after, in this case I want to replace. So I'll create my prompt. And uh I'll go back to that homepage.

Speaker 1: And if I take some text from this uh my blog post, just in case you don't have time to read it, or you have a five-year-old who you want to show it to, then post it here and now use the new prompt um which will take that and rewrite it very simply and uh and here's the answer so The ELI5 might not be useful for you, but hopefully that shows how quickly you could pull together prompts which are relevant for your own use cases. For example, which might be about creating a consistent tone of voice for your content or rewriting content in British English rather than or American English.

Speaker 1: uh or anything else that's that's more specific to you. So there's a lot of flexibility with what is already available in Workshop AI and really encourage you to try out what's there already But there's also a lot coming up soon and the best way to keep on track of that is to follow the Motilai project You can see the roadmap. Actually, these are both private at the moment, but I'm looking at there's no reason for them to be private, so we'll open these up. So you can see some of the things that we're working on now. There's a split between the two back-end projects, so Ragtail AI, but also Ragtail Vector Index. I won't get too much into that now, but that's a Vector indexes are a sort of new type of database which are used very frequently alongside large language models and they're a way of representing

Speaker 1: text but also can be images and other other sorts. of content uh in a way that can helps you identify similarity. So that actually opens up a lot of interesting use cases on the road. But these two things we're working on side by side One of these is uh I can't see the can't see the spike I wanted to talk about. Oh here we go. alt text generation. This is one that we've made some progress on recently. I'm going to show you the quickest way of showing you this is with a loom from my colleague Alex. So he here is adding an image, dragging an image in In the usual way. Chris presses the wand icon and the title of the image is replaced quickly with a beagle dog curled up, lying on a blue blanket. Here's an image which will be familiar to you if you

Speaker 1: use microbakery snowcap mountain. So uh pretty simple, that went rather quickly. That's pretty simple. functionality but something that I think is really nicely integrated in a way that will feel hopefully familiar and natural to people who are working with images already in White Hel and is something that is easy to switch on if you've if you've enabled the Wagtail AI package. So that's a pretty whistle stop tour of where we are with Wagtail and a brief glimpse of what's coming up. With Mikeile, I mean, like I said, it's it's this is a pretty active area of uh progress and development at the moment. And uh as with the with the email sending package somewhere where we are really interested in your feedback and opinions and contributions.

Speaker 1: So I hope you'll firstly try it out. It's really simple to install and start using now. And secondly, let us know how we can make it better. I have seen some questions pop up here. No, no, there's anything.

Speaker 2: There's a couple, Tom. So we have got um, would it be possible to opt into some AI features and not all? That's one question. From Trish?

Speaker 1: Yes, so um first firstly you can if you didn't want to have any of those prompts that I showed you can just remove all the prompts and uh not have them available to your editors. Secondly, on, for example, on the image alt generation, that requires the use of what's referred to as a multimodal model, lots of m M's in there. And but you don't have to use a multimodal model, you could use um uh text generation model and and not supply multimodal model. So at the moment there's I guess two main commercial ones. The uh GPT and Claude and the main open source one is called Lava , double L at the start. But yeah, absolutely you could choose not to use that. And uh actually there's a bit of a question that's happening in in the in the group

Speaker 1: working on at the moment is how flexible we want to make that and it's a trade-off really between Flexibility on one hand and simplicity on the other. But yeah, for sure right now you can choose which bits you want to use.

Speaker 2: I think that should answer Stephen's question as well. Can the prompts be tied to different modals? I think that's what you're talking about is

Speaker 1: Well, not quite. I mean the text prompts can't actually at the moment. So the the expectation is that you would uh the way it's set up at the moment is that you choose a single LM to handle the the the text prompt so that's a level of flexibility that we hadn't anticipated but uh be interested to hear from you Stephen or anyone else if there's a uh if you know some compelling reasons where we should consider that And

Speaker 2: is there a way Wagtail AI could be used to decide page themes, create models based on prompts with specific fields?

Speaker 1: I'm not sure I can quite imagine that, but if I can, then I I think that's that that sounds that sounds sort of within the bounds of what we're working on. Yeah, actually I'm no, I'm not I'm not going to answer that completely because I'm I'm not convinced uh near Aj that I've completely understood your question, but I will I will message you in chat and uh see if I can work it out.