RAGViz: A retrieval-augmented generation diagnostic tool to analyze LLM attention mechanisms on retrieved documents

Student Presenter(s)

Tevin Wang

Abstract or Description

Retrieval-augmented generation (RAG) is a LLM prompting technique that is growing in popularity, but there are insufficient mechanisms for visualizing the effects of retrieved documents from knowledge bases on the actual output of LLMs. We propose RAGViz, a RAG visualization tool that can diagnose the "attentiveness" of the generated tokens on retrieved documents and how the retrieved context might affect the generation. It achieves this using two main functionalities: (1) token and document-level attention visualization and (2) generation comparison when adding/removing documents. The system we use for RAGViz includes AnchorDR as the embedding model, ClueWeb 22-B english documents as the knowledge base, and LLaMa-2 as the LLM. RAGViz comes with a built-in UI that can be used to query our system.

Mentor

Chenyan Xiong