Pitt-GPT-Plus: Multi-Agent AI for Custom Chatbots
University of Pittsburgh researchers have developed a novel multi-agent AI platform to create custom expert chatbots. Pitt-GPT-Plus is designed to work on local servers, can be agnostic in nature, and tailored to match the exact needs, standards, and ethos of an institution. This software can be used to create and validate expert chatbots built on local servers using a large language model (LLM) framework specifically built to serve the institution’s (e.g., hospital) needs. Use of Pitt-GPT-Plus on local servers ensures complete control over data handling of sensitive information (e.g., trade secrets, medical data, protocols), protecting intellectual property, and reducing the costs and security risks associated with external cloud-based alternatives.
Description
AI-powered chatbots are increasingly used to provide tailored conversational, expert advice to end-users. However, current commercially available options are cloud-based requiring transmission of sensitive data externally, raising privacy concerns and posing a threat to intellectual property. Pitt-GPT-Plus is designed to work locally on an institution’s servers, alleviating privacy concerns. Additionally, Pitt-GPT-Pitt can be curated based on an institution’s hardware requirements and authoritative sources of information to provide expert knowledge to end users. This novel approach could revolutionize the use of AI-chatbots in providing expert assistance to users in areas including health, military and other commercial and legally sensitive fields.Applications
- Remote or field medical support chatbots- Military support chatbot
- Technical standard operating procedure chatbot
Advantages
In addition to privacy concerns, current AI-chatbots also have several shortcomings including costs, with usage-based application programing interface (API) costs that can scale unpredictably with adoption or workload spikes. Additionally, many AI-chatbots deliver generic answers not grounded in institution-specific information which can lead to hallucinations and accuracy issues. Finally, many commercially available AI-chatbots cannot be audited, limiting the transparency behind decision making and scoring, limiting validation and deployment.Pitt-GPT-Plus is designed to overcome these shortcomings. Firstly, Pitt-GPT-Plus is designed to be operated locally allowing for complete control over data and costs. Secondly, the system is designed using a specialized open source LLM model, designed to address the institution’s specific data needs, and an advanced multi-agentic retrieval augmented generation framework to ensure the LLM responds based only on authoritative information provided by the institution. Together with a multi-agent interactive component, this approach reduces the risk of hallucination and validates outputs, ensuring end-users are provided with responses in line with institutional knowledge.