LLMs powered by transformer models have taken the world by storm, with products like ChatGPT reaching over a million users within the first month of release and surpassing 180 million users by May 2024. While these technologies are incredibly useful, ensuring their safety and accuracy is paramount.

Given LLMs' vast knowledge and capabilities, enterprises are concerned about the potential misuse of AI-implemented features to generate content that falls outside the scope of business operations or could be harmful, thus damaging the brand’s reputation. Hallucinations, where the AI generates incorrect or nonsensical information, are a critical concern and a significant reason why enterprises are hesitant to fully integrate AI into their workflows and processes.

At ChatGenie, we prioritize safety and accuracy in our Copilot capabilities, specifically in our Chatbot implementation. While foundational models come with built-in prompt guards to ensure safety, we have developed our prompt guard agent as an additional layer to guarantee the safety of our Chatbot capabilities. This agent ensures that our Chatbot can detect and address the following:

• Offensive Content: Filters out profanity, hate speech, sexual content, and harassment.
• Sensitive Information Protection: Protects personal data and maintains confidentiality.
• Misinformation Prevention: Prevents the spread of false information and avoids biases.
• Security: Guards against injection attacks and ensures secure data handling.
• User Experience: Avoids irrelevant, off-topic, overly complex, or repetitive inquiries.
• Legal Compliance: Adheres to regulations like GDPR and CCPA.
• Ethical Considerations: Ensures cultural and emotional sensitivity in responses.

Here’s an animated sample interaction of ChatGenie Prompt Guard Agent in action.

ChatGenie Prompt Guard detects sensitive information. The prompt guard protects the exposure of sensitive information like transaction-related information and personal information.

ChatGenie Prompt Guard detects multiple, off-topic, and complex inquiries. The prompt guard prioritizes relevant and manageable inquiries.

To further improve the quality and maintain the accuracy of our Chatbot responses, we have implemented a response refinement agent that serves as a QA for every response generated by our Chatbot. The design considerations for this refinement agent include:

• Accuracy: Ensuring responses are factually correct and relevant.
• Clarity: Maintaining concise, simple, and unambiguous language.
• Tone and Style: Keeping a consistent, appropriate tone that reflects empathy and professionalism.
• Relevance and Context Awareness: Providing contextually appropriate responses that maintain conversational flow.
• User Satisfaction: Aiming for helpfulness and encouraging user engagement.

To optimize performance, we chose Llama3 for these agents due to its advanced natural language understanding, high-quality language generation, and efficient performance. Its smaller context length compared to GPT-4 suits the simpler tasks of prompt guarding and response refinement, offering faster processing times. Additionally, Llama3 includes robust Filipino language training data, making it effective for localized content. Its advanced filtering capabilities, compliance with data protection regulations, and adherence to ethical standards ensure safe and respectful interactions.

Want to bring ChatGPT-level technology to your Messenger and Instagram interactions? With ChatGenie, you don’t need to worry about complicated flow editors—our AI handles all responses seamlessly. Contact us using the form below to set up a meeting.

‍