Disadvantages of RAG. This is the first part of the RAG… | by Kelvin Lu | Medium

Recently, the rise of large language models (LLMs) has sparked a lot of interest in RAG systems. Many practitioners are eager to learn how RAG can benefit their own organisations, and some businesses have already released RAG-based services. In my previous posts, I addressed my research on how to host and fine-tune a project-specific embedding model[1, 4] and some of the considerations for developing a vector database, which is the cornerstone of the RAG system[1]. In this article, I will explore some of the limitations of RAG systems. If you are unfamiliar with RAG and would like to quickly get an idea of how it works in a case study, please check out[2]. Table of Contents · It Starts With Semantic Search · The Chunk Size and Top-k · World Knowledge · Multi-hop Q&A · Information Loss · Conclusion · References Before we go any further, let's do an experiment. The following code piece compares the cosine similarity score of a query against a series of statements. It uses GCP VertexAI’s

Recently, the rise of large language models (LLMs) has sparked a lot of interest in RAG systems. Many practitioners are eager to learn how RAG can benefit their own organisations, and some businesses have already released RAG-based services. In my previous posts, I addressed my research on how to host and fine-tune a project-specific embedding model[1, 4] and some of the considerations for developing a vector database, which is the cornerstone of the RAG system[1]. In this article, I will explore some of the limitations of RAG systems. If you are unfamiliar with RAG and would like to quickly g