Retrieval augmented generation (RAG) provides the capability to constrain
generative model outputs, and mitigate the possibility of hallucination, by
providing relevant in-context text. The number of tokens a generative large
language model (LLM) can incorporate as context is finite, thus limiting the
volume of knowledge from which to generate an answer. We propose a two-layer
RAG framework for query-focused answer generation and evaluate a
proof-of-concept for this framework in the context of query-focused summary
generation from social media forums, focusing on emerging drug-related
information. The evaluations demonstrate the effectiveness of the two-layer
framework in resource constrained settings to enable researchers in obtaining
near real-time data from users.

检索增强生成 (RAG) 通过提供相关的上下文文本，可以约束生成模型的输出，并减轻产生虚假信息的可能性。我们提出了一个双层 RAG 框架用于查询焦点答案生成，并在社交媒体论坛中对此框架的概念验证进行评估，重点是从新兴的与药物相关信息中生成查询焦点摘要。评估结果表明，在资源受限的环境中，双层框架能够帮助研究人员从用户那里获取几乎实时的数据。