CORAL AI:A Hybrid Chain of Retrieval Augmented Thought Model

Abstract

Coral Reef researchers and policy makers often spend excessive time in manually searching through scientific literature to answer specific questions. While modern AI devices such as DeepSeek can speed up this process, they suffer from hallucinations (incorrect but admirable answers) and rarely provide verification sources. To solve this, we designed Coral AI- a special tool that gives an accurate and reliable answer to coral related questions. Our approach connects the Hybrid Retrieval-Augmented Generation Thought model with a curated database of coral research papers. First, we divided the major literature into semantic chunks using spacy module, converting them into a search vector in FAISS vector DB and using the unit relationship graph to store relationships by using Networkx graph DB, and stored with the appropriate APA 7th version citation as their source metadata. When a user asks a question, the coral AI vector DB and Graph DB find the relevant lesson by using the semantic equality search and recover the top rank reference from both databases. The large language model produces an accurate answer using the model (GPT-4o-mini), while strictly follows the recovered reference. Provides quotes for transparency and verification. To ensure high reliability, we evaluated the performance of Coral AI using context precision, recall, faithfulness and Answer relevancy, similarity and correctness. Additionally, we designed an intuitive interface with flask to make the tool accessible for researchers. To grounding reactions in verified literature and to eliminate AI hallmarks, coral AI brides the gap between rapid information and reliable, citied answers-works more efficiently and confidently to coral researchers and policy makers