Translations:FACTS About Building Retrieval Augmented Generation-based Chatbots/4/en

Enterprise chatbots, powered by generative AI, are rapidly emerging as the most explored initial applications of this technology in the industry, aimed at enhancing employee productivity. Retrieval Augmented Generation (RAG), Large Language Models (LLMs), Langchain/Llamaindex types of LLM orchestration frameworks serve as key technological components in building generative-AI based chatbots. However, building successful enterprise chatbots is not easy. They require meticulous engineering of RAG pipelines. This includes fine-tuning semantic embeddings and LLMs, extracting relevant documents from vector databases, rephrasing queries, reranking results, designing effective prompts, honoring document access controls, providing concise responses, including pertinent references, safeguarding personal information, and building agents to orchestrate all these activities. In this paper, we present a framework for building effective RAG-based chatbots based on our first-hand experience of building three chatbots at NVIDIA: chatbots for IT and HR benefits, company financial earnings, and general enterprise content. Our contributions in this paper are three-fold. First, we introduce our FACTS framework for building enterprise-grade RAG-based chatbots that address the challenges mentioned. FACTS mnemonic refers to the five dimensions that RAG-based chatbots must get right - namely content freshness (F), architectures (A), cost economics of LLMs (C), testing (T), and security (S). Second, we present fifteen control points of RAG pipelines and techniques for optimizing chatbots’ performance at each stage. Finally, we present empirical results from our enterprise data on the accuracy-latency tradeoffs between large LLMs vs small LLMs. To the best of our knowledge, this is the first paper of its kind that provides a holistic view of the factors as well as solutions for building secure enterprise-grade chatbots.