Translations:FACTS About Building Retrieval Augmented Generation-based Chatbots/69/en

Active Retrieval augmented generation (FLARE) (7) iteratively synthesizes a hypothetical next sentence. If the generated sentence contains low-probability tokens, FLARE would use the sentence as the new query for retrieval and regenerate the sentence. Mialon et al. (12) reviews works for advanced augmented generation methods in language model. Self-refine (11) builds an agent to improve the initial answer of RAG through iterative feedback and refinement. ReAct (16) Agent is widely used for handling the complex queries in a recursive manner. On the RAG evaluation front, RAGAS (4) and ARES (14) utilize LLMs as judges and build automatic RAG benchmark to evaluate the RAG system. Zhu et al. (17) overview the intensive usages of LLM in a RAG pipeline including retriever, data generation, rewriter, and reader. We believe that our work provides a unique perspective on building secure enterprise-grade chatbots via our FACTS framework.