Chatbots

RAG Chatbot Implementation

The simplest way to explain a RAG chatbot is picturing a person looking things up in a library before telling you the answer. Sometimes memorized words, sometimes slightly adjusted because you asked for it in their own words. FluffyBot is a mixture of known responses to various prompts/intents and a backdrop of being able to answer questions based on its library of information. You can find FluffyBot by clicking the chat button in the bottom right of any page.

Implementation Details

My previous experience using Voiceflow gave me the groundwork to understand intents, responses, and organize conversation flow. Switching to a more granular tool in AWS allows me to have far more control over what model I use for people to interact with when looking up information, it allows me to create branding, and provides an updated example of chatbot creation. If Claude has a low-cost high-performing equivalent, I could switch to the other model without much trouble.

Creating the chatbot in such a hybrid manner means being able to save costs, by answering commonly asked questions before the request reaches the LLM portion of the bot. We can also adjust usage rates as needed to accomodate website visitors while staying within budget. Updating the chatbot can take time to do properly, or we could simply update the knowledge base with our new documents and press 2 buttons to let our model reference the new information.

I then customized the chat interface for my desired parameters. Have a picture of Fluffy I created playing a game, use a color scheme that matches the site, and have the ability to swap to voice for accessibility. I've also adjusted the bot to only work if running directly from my domain, so attempting to run the bot elsewhere won't work. Other versions can be made with a helpline available, a menu for common actions, and more.

Key Features

• Fully customized layout, interaction, and model used
• Cost saving by incorporating FAQ into registered intent responses instead of solely relying on RAG
• Easy conversation history storage to analyze usage
• Voice capability for accessibility

Design Principles

Voiceflow Chatbot

I decided to learn about Chatbots. The difference between FAQ bots (those that have been around for a long time), AI bots, and Modern bots. While there are many services out there for quickly and easily creating a chatbot, such as Chatbase, Dante, or Cody, I decided to have more control learning and using Voiceflow. They have a surprisingly great degree of options and graphics user interface (GUI). By supplying some of the website pages as the Knowledge Base, I can have the chatbot refer to information throughout the site as it's updated.

For those not aware, AI uses a metric called tokens for calculations. Tokens can best be thought of a 'keywords', 'parts of a keyword', or even a single letter of a keyword. It all depends on how the information was separated. Voiceflow will allow me to use this chatbot up to a limit of tokens that have been processed, and will ask for an upgrade or wait till the following month for a new amount available. So if the bot doesn't appear to be working at the moment, thank you for your patience.

This chatbot can be found in the bottom right of most pages and may be considered a custom-knowledge chatbot. It's meant to serve as an easy way to find information throughout the site contents without needing to be on the correct page detailing it. If the information can't be found, it should offer the whitesheepgamlingllc email as an option for requesting additional details or reporting issues.

While a standard FAQ chatbot is all that's needed, I wanted to explore adding logic for a better AI experience. I started programing supplied text with possible alternative versions, passing responses into variables that may go through a services similar to ChatGPT, and structuring the flow of a conversation that loops back to asking if there's anything else it can help with. It will hopefully pick up on the intent of your statements and direct you to the correct information in a more conversational approach than a simple FAQ. Later, it will be able to provide summaries and links to the most recent updates White Sheep Gaming provides.

[Further Application] These skills can be easily adapted for customizing a bot for business intake. Training the bot on company specific information and have the bot present a fillable form in a conversational approach, which would then be sent to a database. Tools, such as Power Automate, could then take the information from the database and present it in a visually appealing format within an email. You could also summarize totals or send an automatic confirmation, but those are all Power Automate things we can talk about in a later project.