Abstract
Retrieval-Augmented Generation (RAG) has become a dominant paradigm for grounding large language models (LLMs) in external knowledge. However, prevailing RAG systems remain fundamentally text-centric, relying on arbitrary document chunking strategies that often misalign with user intent, obscure provenance, and introduce inefficiencies at query time. This paper introduces FAQ-RAG, a structured alternative that treats questions and answers as the atomic units of knowledge. FAQ-RAG transforms documents into exhaustive sets of frequently asked questions paired with grounded answers, each stored as a dual-vector representation capturing both intent and substance. By indexing knowledge rather than raw text, FAQ-RAG improves retrieval precision, recall, and citation fidelity while reducing reasoning overhead during inference. We present the conceptual framework, system architecture, and practical advantages of FAQ-RAG, and situate it within the broader landscape of retrieval-augmented and citation-aware question answering systems. The source code is available at github.com/Pro-GenAI/FAQ-RAG.
The source code is available at
github.com/Pro-GenAI/FAQ-RAG.
Keywords: Large Language Models, LLMs, Retrieval-Augmented Generation,
Question Answering, Vector Databases, Knowledge Representation, Artificial Intelligence, AI