Translations:FACTS About Building Retrieval Augmented Generation-based Chatbots/22/en

Retrieval Augmented Generation (RAG) is a process where relevant information is retrieved from vector databases through semantic matching and then fed to LLMs for response generation. In a RAG pipeline, vector databases and LLMs collaboratively ensure the delivery of up-to-date enterprise knowledge. However, RAG pipelines have many control points, each of which when not tuned well can lead to lower accuracy, hallucinations, and irrelevant responses by Chatbots. Additionally, document access control permissions complicate the search and retrieval process, requiring careful management to ensure data security and relevance. Furthermore, multi-modal content necessitates the use of multi-modal retrievers to handle structured, unstructured, and semi-structured data, including presentations, diagrams, videos, and meeting recordings. Addressing these challenges is critical for maintaining the accuracy and reliability of enterprise chatbots. Inspired by (3), we identify fifteen control points of RAG from our case studies visualized in Figure 1. Each control point is labeled with a number. In the remainder of this section, we present our insights and learnings for addressing RAG control points.