MEET ARIE

ARIE is more than a chatbot. It represents ASPPH’s commitment to the ethical and responsible use of technology in academic public health. The chatbot is designed with care to:

Unlike traditional AI chatbots that rely solely on generic large language models (LLMs), ARIE utilizes Retrieval-Augmented Generation (RAG) to dynamically incorporate vetted, up-to-date, and domain-specific content.

RAG combines two key steps to improve the accuracy and relevance of responses. First, it retrieves information from a trusted database or knowledge source, like ASPPH’s Academic Resource Hub, a digital library of curated documents. Then, it uses that information to generate a thoughtful and context-aware answer that includes citations to the referenced document.

Selecting the right large language model (LLM) for RAG is also critical because the LLM determines how effectively retrieved information is understood, synthesized, and communicated. For ARIE, ASPPH selected Anthropic’s Claude 3.5 Sonnet V2. Anthropic is a pioneer in the development of Constitutional AI, a method used to train AI systems with guiding principles based on the UN Universal Declaration of Human Rights.

While the RAG framework and underlying LLM provide a solid foundation for improving AI’s accuracy, contextual relevance, and reduction of false information (often referred to as hallucinations), ARIE encompasses other key components that further enhance its ethical and responsible usage.

AI System Prompts

Define the AI chatbot’s purpose, role, and tone. They provide background information and establish the context to help the AI understand the specific task at hand and tailor its responses accordingly. Think of them as ARIE’s “house rules” that guide and influence its behavior and output. The ASPPH project team carefully crafted ARIE’s system prompts specifically for academic public health and continually evaluates and updates them.

 Amazon Bedrock Guardrails

An AWS service that simplifies the safety, privacy, and compliance of AI chatbots. The service provides powerful features for ARIE to filter out harmful or inappropriate content, detect and redact Personally Identifiable Information (PII), and ensure generated responses are supported by underlying source documentation.

Amazon Bedrock Evaluations

An AWS service that brings rigor and clarity to the LLM and RAG evaluation process through customized rubrics and human review workflows. The service ensures that ARIE meets the high standards of an academic environment and delivers the desired performance while aligning responses with ASPPH’s mission and values.