Building a Business Chatbot with GPTBots.ai

In this sponsored October 2023 walkthrough, I demonstrate GPTBots.ai, a service for building business chatbots with language models, knowledge sources, memory, and workflows. GPTBots.ai sponsored the video. I have set up an example for my consulting company, Forward Future, and will show the basic configuration rather than claim that the bot is already proven in production.

In the developer portal I create an AI assistant called FF Bot to answer questions about my services and prices. The interface also offers a knowledge Q&A bot, while a character bot is listed as forthcoming in this version. I set the opening message, optional introduction and tags, and an identity instruction asking the bot to assist visitors helpfully and politely.

The 2023 interface offers model choices including ChatGPT and Claude; the ChatGPT options shown include GPT-3.5 Turbo with 4K or 16K context and GPT-4 with 8K context. I leave the temperature at 0.3. Model availability and names are historical, and the right choice depends on the task and testing.

The knowledge settings allow sources such as URLs and documents. The interface exposes a document-correlation setting, options for unrelated questions, and whether to show knowledge sources in replies. These controls influence retrieval and presentation, but a higher setting does not by itself guarantee a more accurate answer.

Memory settings include long-term information from prior conversations and short-term recent chat history; I select the last five for short-term context. A token-allocation control divides space among memory, identity, knowledge, user input, and the response. In my configuration, the knowledge base gets 40% of the displayed allocation, and I enable suggested follow-up questions.

Under Knowledge, I can add a dataset from text, a URL, a file, or a Q&A document. For this example I supply my company’s homepage, which the service ingests for future answers. The dashboard lets me inspect and refresh the source after a website update and test vector search.

My first test asks what services the company offers. The answer based on the website is not accurate enough, so I use the interface’s conversation-training feature to provide the desired response: education and training, custom AI models, and secure infrastructure. The correction improves this example; one test does not establish that every future answer will be correct or that the underlying model weights were fine-tuned.

The dashboard also has chat records for reviewing conversations after deployment. My bot is not live in this demonstration, so it has no customer records yet. Reviewing real questions and correcting responses could help improve the bot over time.

To put the assistant on a website, the demonstrated interface offers an embed option, a preview, and controls for the color, icon, and avatar. It provides iframe code to add to a page. The service also exposes APIs for conversations, documents, bot information, and speech recognition, so some operations can be handled outside the dashboard.

For a more controlled flow, I show a drag-and-drop Flow Bot. Its input connects to a branch predicate that classifies the question as services, pricing, or other. Each branch routes to a model node and then to an output node; the service branch contains the information I want the bot to use.

I test a services question and watch it pass through the classifier, the services node, and the output. The visual editor can also connect a knowledge source rather than requiring all service details to be typed into a node. Other available components include preset responses and Boolean predicates.

The plugin page includes a Google Search option for adding web search to a bot. In this interface it requires a search engine ID and an API key; it is separate from the uploaded knowledge base. The platform also offers a way to publish a bot to a marketplace for others to use.

At the time of the sponsored video I say that one bot can be created free to get started. Plans, limits, integrations, and the interface may have changed since 2023, so check the current service details before relying on that offer.