This is a step-by-step tutorial (by Speaker A / LARRUZ) demonstrating how to build a business-focused chatbot using a no-code visual builder (sponsored by GPT Bots). The video walks through creating a bot named "Hyper" for a hypothetical company (Hyperscale): choosing a template or blank bot, selecting an LLM (speaker chooses GPT4O), configuring hyperparameters and an 8,000-token context budget, writing an identity prompt/persona, uploading PDFs (or other files / website links / Q&A pairs) to form a knowledge base, selecting OpenAI embeddings (text-embedding-3-small) and chunking rules, testing retrievals and tuning similarity thresholds, disabling/relying on integrations and long-term memory, enabling chat history and optional voice (OpenAI Whisper), deploying via API/link/iFrame/Slack/Discord/WhatsApp/Zapier, testing the embedded bot on a website, and iteratively improving the bot by editing responses to retrain the knowledge base. The speaker highlights team invites/roles and notes some third-party integrations may require payment.
Questions this video answers
August 9, 2024.
GPT Bots.
GPT4O.
8,000 tokens.
text-embedding-3-small (OpenAI embeddings).
Chapters
If you want to add AI to your business applications without writing a single line of code, check out GPT Bots who are the sponsor of today's video. It's a no code visual builder with drag and drop functionality that will let you connect AI to your business applications. It provides a large number of pre built templates that will help you get started in no time. You can also include Agentic workflows which makes for very powerful bots. So let me show you how to get started. Okay, once you sign up you're going to see an interface like this. So you can click on the new bot to start creating a new chatbot for your business application. And you're going to be presented with a number of different templates which have pre built chatbots that you can start using right away. In our case, we're going to be creating a chatbot for a hypothetical business called Hyperscale, which provides computer related services. Now these templates are organized in different categories including marketing, customer support, sales, sales in general, and I'll highly recommend to check them out before you start experimenting. In our case we're going to click on Create from blank. Next you have two different options. One is Bot, which is driven by a single LLM suitable for simple business scenario. And the other one is flow bonds which is helpful if you have multiple LLMs. Think of this as a generic workflow. Now in our case it's a very simple bot, so we're going to select bots and you also want to provide a name. So I'm going to call it Hyper. The bot introduction is optional, but I'm going to say a customer support bot for Hyperscale. Next you're going to be shown this guided step by step process which will let you create your bot. The first thing is you need to select your LLM. Now they provide integration with a number of different API providers including OpenAI, Azure, Anthropic and Mistral. For our case we are going to select GPT4O. You can also set different hyper parameters. So for example, you can control the creativity by changing the temperature. But in my case I'm going to just keep it to the default value. Now in this case the context can only have 8,000 tokens, so you can also control how much of that 8,000 token is going to be used for user input, knowledge base, the identity prompt. We'll talk about this in a bit. Well, as the memory that is going to be assigned to the bot, so we'll just keep everything to default and click on save. Next we need to provide identity prompt. This is basically the part of the prompt that is going to control the Persona of your bot. I took some help from cloudsonnet to come up with this Persona. So here I'm adding character what type of skills this bot is going to have as well as some of the constraints. You can also use the AI to help you write. So they have an internal AI system which can help you iterate on your identity prompt. So I'm going to just use the one recommended by them. Now, since it's a bot or customer support bot for our business application, so you want to make sure that it sticks to our business use case. So for that you can provide a custom knowledge base that is going to be used by the bot to generate responses to the user. So in order to do that, we're going to click on upload. Here you can provide a number of different files. So you can upload word or text files or PDF files, spreadsheets. You can also provide a link to a website, it will scrape the text for you. Or you can add question answer pairs to your knowledge base. Now, in my case, everything is in the form of PDF file. So we are going to do just track that PDF file here. Okay. So right now it's being uploaded. The PDF file is called hyperscale. I'm going to click on next. The next step is to select an embedding model. Currently they only support the OpenAI embedding model, but there are three different options. We will select the text embedding three small for splitting of the text. There are two different options currently available. One is to use max number of tokens that you can define and the second one is to define separator in the text that you're providing. So for our case we will stick to the max number of tokens. We can click on update Preview. This is going to show you what different chunks are going to look like. Now, in some of the cases, the length of the chunks are more than the number of tokens that you have to find, but not by much. Okay, then we're going to save this. Now you can see that it's actually learning. So basically it's creating embeddings for each of the chunk and then adding those to a knowledge base. Okay, and you can actually test it. So you can click on test Retrieval. Here you can provide your question and see what the embedding model is going to return. So for example, here's my question. What are the warranty terms? Now if I click here, it doesn't Generate anything and it has to do the correlation. So you can change this threshold based on which it will determine whether a given text chunk exceed that threshold or not. So if you click now we can see that it return five different chunks. So it could be a good test in which you can see what are the hyper parameters that you want to use. Now, in my case we're solely focusing on the semantic search and we are not using a re ranking model. Okay, so once you're happy with your knowledge base, you can also set the responses if it's not able to generate a response based on the knowledge base. So for example, you can have a preset response which basically will tell it that certain information is not present in the knowledge base. But we're going to rely on the LLM next. This platform gives you the ability to integrate your bot with a number of different external tools, so it can really act as an agent. And there are a lot of different integrations that are available which is pretty comprehensive in my opinion. So you can look at those different categories. Some of them are free, but some of them are third party integrations. So you will actually need to pay them if you want to use these tools. For our bot, we don't need anything. So I'm going to just disable this. You can also add memory, so you can enable or disable the long term memory and also assign different attributes to the user. So we're going to keep it very simple. I'm going to enable chat history, but we're going to disable the long term memory. The next step is to set up how the model is going to behave in the beginning. So you can have a preset message whenever somebody starts the chat. So here you can see in the preview that this is the preset message that is going to be shown to the user. You can also add questions which are suggested questions that are going to be generated by the model when the user asks a question. So if I enable this, you can see that it says try asking me. And here's the suggested question from the model. But in our case we'll just disable this since it's business chatbot, so you can enable human handoff. So let's say the model is not able to generate responses based on your knowledge base. You can let the model choose when to hand it off to a real human customer support agent. And this lets you use something like Intercom or Webhook to do that. It's an extremely powerful feature for real business use cases. Okay, next you can enable File upload from the user. So the user will be able to upload file, but we're going to disable that. And you can enable voice chat through the user interface. So here you can see this voice icon or the mic icon. The user can actually speak to the model for the output. You can also set whether you want only text output or voice output. And we're going to save this and you can test the model and see how good it performs. So we're going to ask it to tell us the location we sent that request. And here it says of course, and here's the location that we provided in the knowledge base. Now you can also click on this button that will actually start speaking.
Of course. Hyperscale is headquartered in San Francisco, California. Our physical store is located at 1234 Tech Avenue, San Francisco, CA 94105.
Okay, this works, but it's just on a platform. How do you make it accessible to the external users? There are a number of different integrations, including an API that you can use in your own applications. You can share a link, you can add it to your webpage using iFrame. There's even integration of Slack, Discord and WhatsApp, as well as you can use it through Zapier, which is pretty powerful. Now, in my case, let's say we want to add this to our own website. So we'll just enable iframe and then click on Manage. So here it will give you an integration code that you can use or integrate into your own website. So I asked Claude to generate HTML code for a simple website and I added this IFRAME here. All right, so I'm going to go to one of the online HTML editor, click on Run. Okay, so here you can see a bot and we can start interacting with the bot. So I'm going to ask it, what is your name? It says, hello, I am Hyper, your AI representative for Hyperscale. Next, I'm going to ask it, what services do you offer? So this is going to be an integration into your own website, just like a simple chatbot. Okay, now if you go back to your dashboard, you're going to see the Hyperbot here. If you click on the bot, you're going to be brought up to the same space which we used to create the bot. Now, if you click on chat history, you will see that it will show you all the different chats that has happened so far and where the chat was coming from. There's a neat feature in here which is action. So you can actually train the bot. So the way it works is you, let's say, look at one of the prompts that the user used and the bot generated a response. And if the response is not correct or something that you don't like, you can change the response and that will be added back to the model's knowledge base. So let's say in this case it spelled the hyperscale wrong and the actual knowledge base, I had the spelling wrong intentionally, so I'm going to just correct that. We also need to provide a name of a new document. So this will basically create a new document for us. We're going to use the text embedding in this case again. Click OK and then you can see that it actually used that information to train the model again. You can also invite other team members to collaboratively create and train bots. So click on this invite link. You can already see that I have invited a couple of other team members. You can assign them specific roles, provide the email addresses, or you can generate a link that can be used to invite team members. Now overall, it's a very powerful platform that lets you build very powerful business driven and business oriented chatbots. I really like the different integrations that it has, so really gives you a complete business solution. It also gives you the flexibility of adding different models so you're not limited to a single API provider for audio input and output. It currently has the OpenAI whisper. If you're building chatbots for a business application, I'll highly recommend to check them out. Link is going to be in the video description. I hope you found this video useful. Thanks for watching and as always, see you in the next one.
