Using ChatGPT with Servoy – Jean Silva, Jan Aleman
Jean Silva and Jan Aleman from Playbookify explain how they are using ChatGPT with Servoy
ChatGPT Integration in Playbookify
- Playbookify introduces a new feature that leverages ChatGPT to automatically generate job descriptions, key responsibilities, and company playbooks.
- ChatGPT offers benefits for content generation but poses challenges in formatting responses for specific use cases.
- A live demonstration showcases how to generate a playbook using ChatGPT, including sending prompts and receiving generated content.
- Crafting prompts carefully is crucial to obtaining the desired output from ChatGPT.
Integrating ChatGPT with Applications
- A step-by-step guide is provided on integrating ChatGPT with applications, covering account creation, API key retrieval, and making API requests.
- Choosing the right text for prompts and refining responses to meet specific requirements are discussed.
OpenAI API and Code
- The presenter demonstrates using the OpenAI API to generate responses with the GPT-3.5 model.
- The code receives input, sends API requests, and returns responses.
- Headless Chrome or batch processors can be used for faster response times.
Fine-tuning and Data Privacy
- Fine-tuning models is possible by creating new models and feeding them cases.
- Data privacy is addressed, clarifying that data is not used for training and can be set for zero retention upon request.
- Training costs are mentioned, and contacting OpenAI for discounts is suggested.
Output Filtering and Prompt Formatting
- Output filtering is not yet implemented but is being experimented with.
- An example of prompt formatting for fine-tuning is provided.
SOP Generation and User Data Rights
- The potential for generating SOPs based on existing ones using fine-tuning is discussed, emphasizing data privacy considerations.
- The user’s data has rights and should not be fed to the AI without explicit consent.
Additional Features and Considerations
- Users can write prompts in J GPD format, such as “how many bananas in one calorie.”
- Generating content in HTML format is possible with specific prompts.
- Additional prompts can be added, like generating a checklist at the end of a Playbook.
- Playbook 5 will introduce more interactive playbooks with AI-generated text boxes, questions, and images.
- Privacy settings apply only to API endpoints, while the web-based version retains information for model training.
- Users can disable data retention in the settings menu, but data will still be retained for 30 days.
- Self-hosted solutions for faster training with user-uploaded data have not been explored.
The Jet GPT into Open Air is open air, it’s Jet GPT. It’s opening air is the company. Jet GPT is open air. It’s the product. He’s going to show you a step by step and you can fill the law with you. Ah, well, they will need a key. So I don’t know. They will need a key. Oh, go to Open Air, I get a key. Then you can fill the law with the job. But you will show the details and we’ll share the code with you on this part. So if you want to do this by yourself, it’s actually pretty easy as people. So you will show you what you figure it out, but you spend a lot of time doing. But before I do that, I will give you a brief overview of why we’ve built this and how you should be looking at this as well probably. So Jet GPT is particularly useful if you need to generate a bunch of content. And with very few hints, it can generate pretty good content, maybe as a starting point, maybe as something useful in your application. So the first thing we came up with is the year and roles. So if you’re playful, you can create all the company, divisions and then the roles within your company, which is fine because that’s typically what you want. If you’re running over it’s all you want a bunch of roles. But now let’s say as the boss, you’re trying to describe the role of marketing manager and today, you know, you have too many beers yesterday, today you’ve got to feel deeply inspired. So then what you did here is when you’re in the field job description, you have a monthly compiler generate. And then based on all the title, it will generate a job description for this function. And then you can go, oh, I need to say, for key responsibilities. And then it will also generate the key responsibility, you don’t have marketing manager. And then based on your company, you can go in now and say, yeah, this makes sense. You’ve got a marketing strategy, yeah. And then you can go in and do the latest, then move it out. But as you can see, this generated a whole bunch of pre useful data in like a second instead of you having to come up with everything. I would play with, we saw the same problem, but a lot of companies realize, yeah, we need SOPs, we need company, handbooks, we need company playbooks. And then you just start at this very high. So that’s why we also filled it in here. We’ve got with a couple of problems that there’s a couple of problems you send to it, not. So based on, you just put in a name of your playbook and then based on that job, send it to a bunch of problems to chat you the team, and it will generate an entire playbook for you. Yeah, well, wait, well, you described it all, I’ll generate mine because it takes a time to go. Okay. So this will run in the background as a server side processor, you can continue your work in the slides, you run in the best processor. So I’ll make a very useful playbook for all of you. This is my promise to you for today. How? Any coverings? This would be a good one. We’ll do this one after, it could be fun. How do you create? Park free software. And then we call it to this. That’s what suits, sir. Come on. You’re nice today. What? Then you click on Generate, and then it will tell you, look at already changed the title of you. So nice running in the background takes about 20 seconds or so. And then we’ll send you here an un-mushification center to come with the multiplication when it’s done. So now you can go to other things. Maybe you can explain what is doing. Yeah, because just sending a completion request to Chachi, open AI, API, ChachiD, it’s fairly easy. That’s why the tricky part is when you want to use ChachiPD in your application to fit your problem. For example, here, I was, the trick is part was figured out how to come up with those questions with ChachiPD to ask them in some sort of format that I could translate into my found sets and create content behind the scenes. So this is what you have to be focused on. That’s the trick is part. And the rest is just the IP call. Use pre-p3 times. Yeah. Yeah. You want to send out you? Yeah. Somebody has to do it. Yeah. And is you going to show it in a second? So let’s look at the, is it done yet? That’s how I look good. Yes, it is done. So that’s your playbook. As you can see, it does some formatting. You know, all the bullets and formatting remains intact. Understanding colon software bugs, implementing effective software testing strategies. That’s pretty good. Debugging. Debugging. Debugging. Debugging. Debugging. Debugging. Debugging. To be produced. Defining conquer. Fineracing. First split, oh, this is brilliant. So as you can see, you know, you don’t have to think anymore today. You just go ChachiPD and it does it all for you. Just go ChachiPD. The I will show this label on the, on the discord channel. By the way, by the way, by the basis is free to use. So don’t keep your sharing and creating an account. You can use it for free. That’s my offer to you today. Let’s head out of her way. I’m going to give it to John how to actually build this. John. Yeah. Yeah. So first, I think why everyone already understands why we need to go. You have to use those to see how we are on a new way or now in things are constantly changing and you are an AI seems like being a thing now. And yeah, you can use in better cases like, but you find for example, it was one and you probably have one good fit in your home money to use it. So. So the first thing we need to, of course, I probably do that. Case that we feed and you need to generate a create an account and open AI. I put in this slide so share with you guys later. If you follow this, then you will be able to do it. I won’t get the code. And I’ve added key and this is where you can create one. And with this key, you can make requests to the API and you have a lot of and points you can use it. You want that that I’m using in here, particularly is the completion on that. You can ask some chatty team and you are responding back. So basically, we will post requests to this endpoint right there. We’ve asked the key to the better token. And yeah, one of the things that I saw, that I saw, is that we need to choose the text that we’re going to try to get. Carefully, because, well, you can first, I was just trying some combinations to see what was better for me. Based on the algorithm that I created to do the way that I come up to get the rights. So you have this request button that you’re going to send in that request. I don’t know if you guys can. Oh, yeah, that’s correct. So you have this model I was using the chatty API file. And yeah, this right here, this part, it’s just to tell you, hey, to direct the name more into the matter. So, hey, you are now, especially listening right in the description of jobs. Just a little bit. And then later you can, well, generate a description for it. And then this part will happen with the output of the user. And then he’s most back. His most back looks like this. So you go deeper into choices, and then you get into message and content. And then you are good to go. Okay. So how we do that, I’m just going to go these by piece and then go over here. So I can show you guys how I did that. So here I created this scenario where you have a bank of jobs. And you can, based on the title, it’s going to suggest you the job description. And this one was playing yesterday. I’ll simply value operating generates a bunch of things to me here. Since the people. Good. So let’s see how it goes. So he generates this. And of course, it’s not perfect. I didn’t refine it. I didn’t put, didn’t try all the combinations that I could like. Give me a description with this data net or give some examples to me first. Just to refine a little bit better the response, but you can do it as well. I can see that. All right. I created this form here. I like to look into exposure to the view. It’s better perspective for everything. So, in generator description, I’m going to go from the top. I have the function. And in this function, basically. I’m going to go from the bottom. So, I created this scope called OpenAI. And in that, you can put the calls and requests that went to OpenAI. And I created this application that I was showing you guys. And I just passed the question. This is my question. In Playbook 5, I created my question as B. I refined the, like, give me this with Dixie Strictor. No more than any words or something like it. And then this is following a little bit of the guidelines that I’ve learned when you guys later. Because you can also do, like, you can also do is training your model. Creating your own model based on what the model is that you have. And you can feed some information before prior to get to ask something to me. To this model. But the first thing that they encourage you to do is try those kind of manipulation to questions. So you improve your prompt and you get better responses. So here, the code. Don’t get that. So you have this token that you’re going to need. I’ll do it later. So you have this token that you’ve got to get. That’s not linked. I said you guys have to create an account with no AI. So then you create this token there. And you can use it for consuming the API. For this, this is the base URL. The base URL should be something like this. But for this problem, I just, the small like the endpoint that I want to reach. But this is it. Well, you can change it later. This is the model that I’m using. It’s the 3.5 toggle. I tried to make some things with the four. I wasn’t out yet. The equals out for some companies. I don’t know how to get that. Because I tried to get all the code that we then did. Never responded. So yeah, I stuck with the 3.5 toggle. Pergo, G of the, they make the four code. It is. Right. Yeah. Right. Yeah. That was a couple months ago. There was a key to. Yeah. And it doesn’t appear in the, in the, in the, in the, pattern. Yeah. Yeah. I mean, we have this, this Gave Initiative. Thismatically received a message of message of the parameter. And this is the property request that I was showing to you guys, so this can probably, probably can, it can work if I remove this. But this is just to refine a little bit better. Yeah. a little bit more. And then this is the, this is called just, well you guys, I think you guys are familiar with it. And we have, we had a session also about this. So you create the, the, the, the, the clients, make the request, fill in the headers, set the body, and we send the request. And this is a, a secret request. This is just the defensive code. I like to do that. Because the, the responses in it, it’s in choice, the first one, message and comment. And then there you go. You return the, the chat response and then you, you get to go. Right? That’s no hard part of this. That’s all you, you get from the code. And here I just, do some try catch. So yeah, this is just the basic. I’ll share, like I said, I’ll share the code with you guys. And I’ll go back in here to discuss a little bit more about this. So, okay, I’ll show you the code. You now have, have a way to do it. So what now? There are some ideas that might fit by your gaze or what, what, what can do. It’s, use a secret request. That would depend on what, what, what you want to do if you want the user to, because those, those requests, they can take a little bit of time to, to go back with something because, uh, chat with these, not, their API is not to, uh, stop fast. Um, but yeah, depending on how much processing you are, you are doing in your, um, in, in your case, then you can probably go to a head of the spikes like we didn’t do to fly or, or a batch process, or a processor, then that depends on what you can do. What, what you want to do. Right? Because, um, for us, it’s, it’s, um, it’s where we need, we, we, we can, we can, we can feel something and then, well, last wait for the result, or you could also, um, do this in, with batch processing when you’re, um, application starts up and then you start to pick up, um, the jobs that you have in there for processing. And about fine tuning your model, um, is, is it possible to fine tune you, you have, uh, our way to do it? Uh, basically what you do, you create another model and in this model, you, you feed cases into it and create this new model you can use it. Um, like I said before you go down this path, you should like, in looking into, um, how to get, uh, to feed better products into your job, to get, to get good results. So, yeah, I also put some, some guidelines here. If you guys are more interested in, you can, you can look into it and in this example, I don’t know how, uh, how you feed, uh, your model and you create new one. So, yeah, um, I’m dissipating this because I, I know that you guys have this question about the privacy, literally, I’ll, uh, open for, for, uh, discussion with you guys. Um, yeah, what happens with the data that we feed, uh, the charge you, uh, this data, um, when, when you’re talking about safety, I did the, the, the, uh, the build that we were using. Um, every day I just not used to train their model. I, uh, get this from the box, from their dogs. He’s not there. You guys can, uh, really, really, really, more about it than you can not focus and think. Um, however, they, they keep your data for 30 days. Probably for, um, when they need it, they will look into it for, um, for privacy issues or something like it. If you have a fancy forge or something, then some kind of, uh, investigation, police or something. Um, yeah, this, this data will not be used, uh, to, to be training and it’s not, gonna become public. So, okay. I think that’s, that’s why. And, um, yeah. And also, well, I, I told you, well, that, that there is this, this retention thing about this data, but you can, uh, at this end point that I was just using now it’s eligible for, for zero retention. So you might reach out to them and I say, I don’t want my data in your servers. So they might open, um, open a, uh, a, a, a section for you. So those are the networks. And you, you have like, whether this is the one that we are using, no data training and the default retention is 30 days and it’s logical for zero retention. So you probably, that could be handy for, uh, for one, for your project or something. So yeah, I understand why don’t my data in, in 20 other service. Well, actually, this is one. One that I was using, I was just saying. So you have a lot of, uh, inputs you have completion generation that is an, an even audio stuff that you can use out your, uh, our translations. And, um, yeah, that’s, uh, pretty much about it. So I knew, I knew that you guys, who, who, who hope you about this, because that was one of the things that we were discussing, um, during those days. Um, I just, um, for training the data privacy is clear now. How, how expensive is it? If I, is asking, um, how expensive is to train in your models, that’s correct? Exactly. I mean, I’ve data from, um, project on the customer and to have asked about the project, I would need to feed the data from the project or the system to get sensible answers. Well, one thing I said, I see, I saw it’s that they, uh, they charge you for tokens. And those tokens are, um, for example, when you, when you ask for it for data, you see, when you charge it, you, you, you, response like you chunks of, well, like, we’re stipend. If you answer what’s about 400 tokens, the cost and I mean, tokens aren’t very expensive, but training would be the question if you don’t play that there. Yeah, but that’s, I don’t know. Yeah, you can do that. You can get up or, yeah, maybe you can get a, a, a, a, a, I don’t know, trying to get rich salespeople there. Yeah, I think, um, too early, but guys, I want to ask some more questions about something. Maybe, uh, explain some part of the code, I don’t think so, but, uh, it was really fine to get, but, uh, about, about, let us, your, or how we use JIT ideas or something. We have two questions. Uh, sorry. You went first. Do you filter the output in any way when someone is generating a new play, but to make sure there is no strange results coming out of it? Uh, actually not yet. That, that, that’s, that’s pretty experimental. We are explaining the things and see how how we go. But, um, we have, uh, we have, uh, I’ll be do with, if I, that’s, um, yeah, it’s, most of you looking into it. So if there is something, uh, weird there, is going to, uh, yeah, probably delete it or, or, he, he have access to, to those two links to see this in there. So something like breakfast, words. Yeah. Nothing more advanced beyond that. Yeah, not, not ultimately like, yeah, not yet. Can you hear me? How you train your life? Uh, I think, I think, do it fine. You were saying the playbook, if I, you do that. Can you give us an example? Uh, yeah. So I did that, uh, just by questions, um, uh, formatting well the, the, uh, the prompt that I’m sending to chat to you for example, um, the text should not be no longer there than the three paragraphs or something. And I go section by section. So I said, well, that’s more fine tune and then training. I think the question is, well, how do you think, check it all your data? We don’t do that at all. So, So, that’s my question is your finding your life’s evening. So can you, show us an example of what you sent as a fine tune? Yeah. You’re like, this is a good thing. So, for playbook, fine, this is a very early implementation. Yes. We, we do think that, you know, once we have live or customers who have like hundreds of S.O.B.s and playbook, the paper go in and then for themselves generate more S.O.B.s. based on existing ones. Okay. Even do that. So you could say, this is how our company writes S.O.B.s. that new people coming into the company could use that as a starting point. But for that, we need to think about data privacy as well. Because as a playbook, if we are, if we have access to thousands of accounts of other customers. And obviously we can just take their data and send it over to Jetty Petiters. Data ownership issues there. So we need to get explicit consent. So that’s something that you guys should keep in mind as well. If you’re building applications for other people, you know, their data has rights as well. So you have to do it. Mind you, you just can’t feed everything to the animal. Let us Jetty. Your off screen still. Yes. Oh, okay. Oh, your ex on purpose. Yes. Yes. Is hacking some code together? We don’t want to see those images. No, no. Now we don’t want to see them. We definitely don’t want to see them. So the question is, have you written your prompt yet? Have you written your favorite prompt yet to Jetty Petit? Something with bananas maybe? Yes. How many bananas in one calorie? Okay. We can probably do this. Hang on. That’s my fault. Yes. Thank you. Thank you. How many bananas in one calorie? The number of bananas in one calorie can vary depending on the size of the banana. On average, a medium-sized banana contains 105 to 110 calories. So if you’re looking at a single calorie, it’s just a fraction of a banana, like a small slice. It’s pretty good. For example, so just so we can give you a concrete example here. We have a lot of chapters in play with the fire. So I cannot go to the Chachi PD. Hey, give me a holy book and then I’ll get all this structure and create atoms in any play with the fire. Well, I’ll put it, but it’s complicated because those responses are never the same. So that’s the thing that I invited home. So I will for each chapter and I will say this. Hey, I’m already writing a playbook about this matter. Please write about this chapter and write this in French Portuguese because that will depend. This is also including translations. I’m not getting the data in a translator. I was just getting in the language already. So I give me the content in HTML format because this I can understand. I can process in breaking two pieces and feed my database with it. So this is just an example of a little bit more elaboration. And this is just one prompt of others in there. So here’s an example of a playbook that would make sense of the end to run a checklist. Yes, you could tell it. Give me a checklist of all items. I should be checking at the end of this playbook. And it’s pretty good in generating that for you. Yeah. And this is a part that we are experimenting for the next feature in play. We can also provide the way for play groups. We can now because now you see that it’s just content. But we are going to also add some text books, some questions. And I’ll provide it by AI. So of course you can change it and shape it up the way you want. But it’s just to give you some ideas. And you can do the same with images like generate an attractive image for this specific playbook. Because now in playbook, if I do have a placeholder for an image. And your SOP is not looking much more attractive if you can have a generate an image that is the subject. But that’s still the image generation is still very early. When I created one for handling complaints in the hotel, it was pretty racist. It was pretty racist. So for images I think we’re getting close, but we’re not there yet. They need to find you at the bottom. Yeah, I think if you guys have some more questions. Just one thing point out is people, the privacy stuff that you put up, OER applies to the API and points. If you go to the web-based version of it and stuff, I think stuff, it does retain that information and it does use it to train them at all. So it’s OER, the API that has privacy. Yeah, that’s that you need to take care. When you are there doing this, I’m feeling cold to this, to the chat. But you have you have a way to turn it off. Now I saw this there yesterday as well. So you go in there and you insert this in your house, this button there. And you turn it off. I’ve already needed mine. So, but the only thing is that you have to keep in mind as long as that thing about the 30 days, is you have your data for 30 days. Even with this turn off, then you have to take care of what you feed in there. But it can turn it off now. So you need to cannot be used to train models. More questions. You get more points for anything. There’s one more question here. The question would be did you experiment with self-nourless solutions for AI? There are some images in the chat. Which I can basically train faster with data. When we use the uploads information, so I already have information about our projects. Yeah, no, we can end the next one. Thank you.