How to Build Your Own Custom GPT (Step-by-Step)

Full video transcript

Hey everyone, this is Eli with CustomGPT. We’re the leading no-code AI agent builder platform for business, and today I’m going to walk you through a simple use case of ticket deflection and customer engagement. We’re going to build an AI agent here based on CustomGPT’s website, our YouTube channel, our read me, and even a few files using our Google Drive integration.

To kick this off, I’m going to go ahead and click create project. Now from this point I’m going to go ahead and select my first data point, which is going to be our website. Once I paste that in here, our tool is going to go ahead and find the site map for us, and I’ll go and click create project. I do have a few other data sources I want to add in here and allow them to process while we go ahead and customize the settings and everything else.

One other data source I’m going to add here is our YouTube channel. We found that YouTube is a really awesome way to add in extra data, especially for people that may be more visual learners who want to be able to query then navigate into maybe a YouTube video or demo for them. Another thing I’m going to add here is our read me, and then last up I do have a few files from my Google Drive I’m going to go ahead and pull in here — select all.

So now that our data sources are all set, I’m going to go ahead to our agent settings. This is where we have settings on conversation, citations, AI intelligence, some advanced stuff, and also our security tab. To begin, you’re going to select your agent language — this should be the language that most of your users speak in or write in, and also the language that most of your content is in. So here we’re going to go ahead with English.

Next up, if you want to change a placeholder prompt, definitely do that. For this purpose, it’s going to be asking anything about CustomGPT.ai so people know exactly what this agent is all about and what kind of data is in here. A lot of times people like adding starter questions, so we’re going to go ahead with “what is your pricing” to make sure that people maybe have a good conversation starter. Often if someone hasn’t used or queried an AI system or AI agent before, they may not know where to start, and it’s always good to give them example questions which they can go ahead and click on and then get their responses directly from there.

Custom message ending is also a really good way to add in just some extra upside with your AI agent. Sometimes it gives responses and that’s cool, maybe answers someone’s question, but you may have a specific CTA that you actually want to promote — in this case it’s a free trial. For others it might be to put “call,” “contact support,” “fill out a form,” whatever it might be. So we go ahead and click save settings here.

Now I’m going to go back and check on this actual agent. We’ve gone ahead and we’ve seen that all the items are indexed already. If I go ahead and click into here, we’ll see we have our PDFs, our YouTube, as well as any links that we just uploaded. So now I know that it’s fine to go ahead and actually ask questions. Here I’m going to click the example prompt — right now we’re going to see that our pricing, we’ve got our standard plan, premium, and also our enterprise. Additionally, you’ll be able to go and see our actual pricing here, click into it, and it’ll bring you directly to the page that I pulled from. Give a thumbs up, seemingly a good response.

Next up, we’re going to move into the citations section. Citations are an essential component of the CustomGPT system — as you can see here we have an index citation as well as that link at the end there, and you can select how you’d like it to display. What’s really important in our system, and right here I’ll touch on it in the response — source is that our default will only pull from your content. So I uploaded our website, our help desk, YouTube, as well as a few files. Whenever I’m querying this agent, it’s only pulling from those source data sources. It will not go to ChatGPT, won’t go to any other LLM that you might select — it’s just from your sources, and this is called our anti-hallucination algorithm, meaning your data responses come from your data only, not from outside sources. Next up, if you do want to customize some of this stuff like CustomGPT sources, you can go ahead and do that. I’ll save it here.

Moving on right into our AI intelligence area — this is a newer feature that we offer. We have some simple presets for people, so if they want a faster response we’ll switch it over to a faster mini option. Optimal is our more balanced option, and then advanced reasoning uses some more complex machine learning algorithms in the background for harder tasks and questions that require a bit more reasoning or high-level focus for the system itself.

Next up is the advanced section. Here you can enable or disable user feedback — thumbs up, thumbs down. We recently just added conversation sharing and exporting, so if you have a really good conversation on any research topic, for example, you can go ahead and send that to your colleague, send it to a customer, client, whoever it might be. Additionally, if you want to go ahead and remove branding on our premium plan and above, you can do that.

Up is our security area — as I touched upon before, anti-hallucination is amongst the most important features. I’ll reiterate what this means: our system does not make up facts. It pulls from your data, not from the LLM or the wider internet. If it doesn’t have data to answer a question, it simply won’t answer it. So here I’ll go ahead and test this out live for you, and you’ll see right here — because our website, all the data sources, do not have anything about the Celtics — it says “I’m sorry, I don’t know the answer.” You can actually set that preset to whatever you’d like.

Next up, on data protection, we are SOC 2 Type 2 certified and GDPR compliant. There’s no training on any data with our system — we’re a RAG-based system, retrieval-augmented generation, so we’re sitting on top of the LLM that’s chosen. Essentially all the data is encrypted in transit and at rest — no training, simply retrieving from your data set to give you responses using LLMs, not training LLMs or any other types of models. There’s a reCAPTCHA toggle too, of course, if you don’t want to get spammed if you deploy this on a website. You can also whitelist your domain, in the event that you’re putting this on your website and you want some extra protection, to make sure that no one else can pull the embed code directly off the site and just keep it extra secure within your own domain.

Moving on, we’re going to go ahead and build the agent persona. If you have built like an OpenAI GPT, it’s fairly similar to custom instructions here. It’s really important, especially when you’re building something that’s customer-facing, to make sure that it’s up to date and following the rules that you’d like for it to follow, as this will be directly interacting with customers or clients, and you want to make sure the responses in the tone are correct.

Here, for the sake of time, I am going to use our persona generator, so I’m going to tell it that I’m building an agent for CustomGPT’s support function, and this tool is generally a great way to just kick this off — this is all natural language, so you could simply just type in here in normal paragraphs, but I do enjoy this format, especially if you’re not experienced with prompt engineering — it’s quite simple. I’m going to go ahead and click save changes. You can always test out the persona to the right, so if you have specific instructions that you want to try out, for example, you can always save it there to make sure that the way our system processes your change and prompting is in line with what you like.

Next up, I’m going to move to the appearance section — this is where you can really customize how you want it to look. If you’re putting this on your website, you want to make sure it matches your brand, background colors, and everything like that. So I’m going to go ahead into here and grab a few different things — I’ll make our CustomGPT logo, put a background, maybe change this, make it clean, move that there. Great, okay, so now I made it look exactly how I like.

Moving into the advanced section, if you want to change this you can just give it another title or ability there, and you can also add your terms of service — that’s generally important so people know what the terms of service are when they’re interacting with an AI agent within a website.

Next up is the “ask me anything” area — once you have actually done your settings, uploaded your data, and everything, you can come in here and just test out the system itself. This is your sandbox where you can test out, create new conversations, and also see any conversations that were happening. If you do deploy this on a website in front of anonymous end users, they will show up as anonymous here — we’re not pulling up direct names — but if someone within your team account inside CustomGPT asks a question, it will show that they’re the person that queried, and you can go ahead and actually view their conversations right within here.

Lastly, we do have an analytics section — this is just to give you a good idea on overall queries, query volume, user feedback like thumbs up and thumbs down, and then also user locations of the end users.

Last up is the deploy section — a really important component of our system is the ability to deploy this and put it in front of any end user that you want, whether you want to put it on your website, inside internet, on your CRM, deploy and embed on Shopify, Slack, your own SharePoint — you can do that with any of our options and they’re generally no code. This is the most coding you may have to do to actually deploy our system — it’s simply just a copy and paste of HTML code. You can put this into pretty much any CMS provider that you may think of. We do have a bunch of different deployment options — I can show you what these look like in just a second. We have an embed, a live chat, an embed inside a keyword search, as well as a website called Posley.

Everything I’ve showed you today is our no-code app and interface. If you are a coder, or you have devs on your team or something, and you want to do something more advanced, we do have a full functioning RAG API, so all the stuff I just showed you, you can do via API and you can even build out your own front end or implement it into some kind of workflow if you’d like to do that.

Here now I’ll show you what these look like actually fully deployed on our website. Right here you’ll see we use our live chat — this is our modal popup setup. We found that this is a really good way to make it more in a user’s face and engaging. As some of the deployments, if it’s just in the bottom right, people think it’s the general old mill chat bot, click a button and it brings you to support. Instead, this setup actually shows people that it’s more of a full ChatGPT AI-genie type agent that can answer way more complex questions and have more engaging conversations, to a point where you’re really researching, and like reading and conversing with this tool, not just asking “hey how do I contact someone for help.”

Another way that you can deploy this that we have is just a regular embed, and this is a good way again to just put it in someone’s face — it’s a really popular option for many help desks, as people are coming in here they definitely want help, and a big advantage of this tool is someone might come in here and they’ll have a question, for example “what can I do with CustomGPT AI agent index with my data” — right, instead of just contacting the help desk with that, they came in here, asked this question, and get a simple response.

Now one other feature that I had forgotten to mention, and I somewhat breezed over earlier — once I got to these docs I was starting to think, hey, we have all these different guides and whatnot, and we obviously add a lot of new features as we’re a constantly growing company — changes and active features, or make new features. So when we’re adding new guides, of course, when I’m in here clicking around I can go ahead and access it, but what happens directly inside our AI agent to make sure that agent is up to date with the data in our read me? Here I’m going to go back into our sources section, we’re going to go directly into the actions and settings. This feature I’m about to show you here is our auto-sync for sitemaps, and we do also have this for some of our data integrations. What I’m showing you here is when we add new pages or remove pages and update on change, every single week our system is going to go through and refresh and process this sitemap for the read me to add in any new guides, new updates, or delete any old guides or old pages that we had deleted.

What this means is that once you sync this and go ahead and click save changes, I’ll never have to touch the sitemap again. We can go ahead and edit our read me, and then every week — or you can even set it daily if you want to — this system will go ahead and pull in and update all the data, so our AI agent is always up to date, no manual work needed. You don’t even need to think about it — set this, set and forget, and your data is up to date, and your customers of course will be able to access that information and be happy.

All right, and that concludes our demo for today of CustomGPT. Again, we just built a ticket deflection and customer engagement AI agent with no code in just over 10 minutes or so, and it’s ready to go, ready to deploy live on a website. Thank you for watching this — if you have any questions please feel free to reach out at CustomGPT, and also go ahead, talk to our AI agent, and learn a bit more about what we do.