Speaker:Mohammad Ahnaf Sadat

CoralAI: A Retrieval-Augmented Generation Model for Coral-Related Queries

Coral reef researchers traditionally spend significant time manually searching for answers to specific queries. This process has been accelerated by advancements in large language models (LLMs) like ChatGPT. However, these models sometimes generate plausible but incorrect or artificially generated information, a phenomenon known as hallucination, and often do not disclose the sources of retrieved information. To address these challenges, we developed CoralAI, a tool designed to provide reliable answers to coral-related queries. We began by curating a list of research papers and segmenting them into smaller chunks. Each chunk was then vectorized based on its contextual meaning and stored in a vector database with proper APA 7th citations. For each user query, CoralAI generates a corresponding embedding to identify and retrieve similar chunks within the database. Using prompt engineering, we leverage the capabilities of LLMs to provide precise responses directly based on these chunks’ context, and we include the APA 7th citations of these used chunks in the final response. This Retrieval-Augmented Generation (RAG) approach ensures that answers are not only accurate and relevant but also verifiably sourced. We assessed the system's effectiveness through metrics such as Context Recall, Context Precision, Faithfulness, Answer Similarity, and Answer Correctness. Additionally, we designed a user-friendly interface to facilitate easy interactions with CoralAI.

Committee: Simanta Mitra (major professor) and Gurpur Prabhu