Discourse2Draft: an AI-powered structured document writing tool

Introduction We introduce Discourse2Draft, a model-agnostic, AI-powered, advanced Retrieval-Augmented Generation (RAG) based application that converts a user-provided query on a topic into a polished full-form draft written in user-provided outline, using context from user-provided documents or literature downloaded from PubMed or both. This application can be used to generate any document that needs a structured outline. Methods Discourse2Draft can take the outline of a document directly from the user or generate one from the user-provided query using AI. With a structured outline (AI-generated or user-provided) at its disposal, it first extracts the hierarchical structure from the outline and detects positions to write content. It then uses AI to extract core ideas from the corresponding section headers and previously written text for every writable position and to fill those positions with cohesive prose with context taken from user-provided documents or literature. During content generation, it automatically weaves source-anchored citations (user-provided documents or literature) throughout the text. The generated draft with associated references and the bibliography can be exported in Microsoft Word, LaTeX and Markdown formats, ensuring full transparency and downstream editability. Results Discourse2Draft achieved competitive scores against state-of-the proprietary content writer tools while evaluated by LLM-as-a-judge approach in terms of content generation, context retrieval and citations. The evaluation scores were generated and evaluated in multiple runs by sensibly selected generator and evaluator model sets to avoid judgment bias as much as possible. Discussion Discourse2Draft serves the need of an open-source draft writer maintaining reference authenticity, full control on outline management and content reproducibility. By coupling LLM-agnostic evidence-based content generation with traceable knowledge retrieval, Discourse2Draft streamlines the journey from informal textual query to publication-quality writing while keeping scientific rigor and author oversight at the forefront.