Project Overview

The AI Support Chatbot represents a significant advancement in customer support automation for Witherhosting LLC. This intelligent system leverages OpenAI's cutting-edge language models to provide instant, accurate responses to customer inquiries.

We were the first to implement this in the hosting industry and was initially developed before OpenAI released their own version of memories.

Key Features

• Dynamic Memory Management: The chatbot continuously updates its knowledge base from both our helpdesk knowledge base and fine-tuned knowledge injections, ensuring responses stay current with our evolving services.
• Context-Aware Responses: Advanced conversation tracking maintains context across multiple exchanges, providing personalized assistance. We use a state machine to track the conversation and provide a more natural response. We track memory based on long-term memory and short-term memory to reduce the amount of tokens used.
• Modular Architecture: The system is built to attach to any programmable platform, such as Crisp, Discord, or any other social platform.
• Defensive Measures: Multiple subsystems such as rate limiting, text filtering and jailbreaking detection are in place to ensure the chatbot is not abused.

Technical Implementation

The system architecture consists of several key components working in harmony:

At its core, the chatbot is structured around a Finite State Machine, which is used to track the conversation and provide a more natural response. This allows us to provide responses to individuals that tend to send multiple messages in quick succession.

Once the state machine satisfies the conditions for a response, the chatbot follows a series of engines to provide a response. Each engine is a modular component that has their own responsibilities which are executed in a specific order:

• /: This engine is responsible for handling command feedback, usually prefixed by [/].
• A: Handles pre-baked responses that are stored in the local database.
• B: Handles responses that are matched from the Withertron (Discord) keyword database.
• M: Handles the moderation filter for both the chatbot and the user as part of OpenAI's policy.
• C: Checks whether the question/phrase has been asked before, and uses a cached response if available.
• D{...}: Generates a response using the OpenAI API. {...} denotes the version of the engine, due to model changes or internal functionality changes.

Performance & Results

Since deployment, the Crisp Chatbot AI has:

• Reduced average initial response time from 7 hours to under 30 seconds
• Handled over 1000 daily customer inquiries automatically
• Decreased support ticket volume by 72%

The project continues to evolve with ongoing improvements to response accuracy and expanded integration capabilities. Future improvement includes direct integration to the game panel allowing the AI to perform power actions and basic configuration changes for common issues.