In this post, the author discusses the design choices and future work directions for question answering systems that utilize LLMs (Language Model Models) over unstructured data in enterprises, referred to as RAG-U systems. The author explores the options for additional data in prompts, such as chunks of text, full documents, or extracted triples from documents, and the ways to store and fetch this additional data, including the use of vector indices. The post also introduces the concept of using knowledge graphs (KGs) to link chunks of text and identifies potential future work in KG construction, matrix embeddings, and evaluating RAG-U systems.
https://kuzudb.com/docusaurus/blog/llms-graphs-part-2/