Jinhu Qi

Jinhu Qi

Jinhu Qi is a visiting researcher at the HUMANS LAB at the University of Southern California, under the mentorship of Professor Emilio Ferrara. His research centers on Large Language Models (LLMs) applications, particularly in Retrieval-Augmented Generation (RAG) using vector databases and knowledge graphs. Jinhu also explores Chain of Thought (CoT) reasoning, language translation, and misinformation detection. With a Master's in Analytics from USC and a Bachelor's in Computer and Information Science from the University of Oregon, he is currently pursuing a PhD for 2025 to further advance his research in these areas.

Vector-RAG: Optimizing Hallucination Issues in LLMs through Retrieval Augmented Generation

WorkFlow

Building on our previous work, we explored the hallucination problem in LLMs by utilizing Retrieval-Augmented Generation (RAG), a technique that retrieves relevant information from a database before generating responses. I collected data on tourist attractions in Tibet, converting it into vectors using pre-trained BERT and TF-IDF models, which were stored in an FAISS vector database. User queries were also converted into vectors, and we tested various retrieval methods to search the database.

We achieved 65% accuracy in retrieving the top three relevant results. The retrieved information was passed to the LLM, improving its quality score from 0.7429 to 0.9225 using RAG. Our results were published in AIPR 2024, demonstrating that Vector-RAG effectively addressed hallucination and personalization challenges in LLM-based tourism recommendation systems. You can find the paper here.