It begins with structure, either accepting an outline supplied directly by the user or generating one automatically from the user's query, and then fills that skeleton section by section with prose drawn from evidence the user controls. Technically, the system operates in several stages. Once an outline is in hand, the software parses it into a hierarchical representation and identifies every position where content should be written.

For each such position, it uses AI to extract the core ideas implied by the corresponding section header, taking into account the text that has already been written in preceding sections. This sequential awareness is what allows the final draft to read as cohesive discourse rather than a patchwork of independently generated paragraphs, since each new passage is conditioned on both its assigned topic and the surrounding narrative context. The evidence layer is where retrieval-augmented generation does its work.

Context for each section can come from documents the user provides, from literature automatically downloaded from PubMed, or from both sources combined. As the draft is generated, the application weaves source-anchored citations throughout the text, so every claim can be traced back to the document or published paper that supports it. A bibliography is assembled alongside the draft, and the whole package, text plus references, can be exported in Microsoft Word, LaTeX, and Markdown formats for downstream editing.

A notable design decision is that the system is model-agnostic. Rather than being locked to a single large language model, Discourse2Draft couples LLM-agnostic content generation with traceable knowledge retrieval, meaning the underlying generative engine can in principle be swapped without changing the surrounding workflow. The intended scope is broad.

The emphasis on user-provided outlines and user-provided documents reflects a deliberate philosophy: the human author defines the argument's shape and the evidence base, while the AI handles the labor-intensive assembly of prose within those constraints. Transparency is a recurring theme in the tool's design. The export options in standard formats ensure that the AI's contribution remains an editable starting point rather than a black-box final product, keeping the author in control of the finished text.

The article, received in July 2026 and accepted in October 2026 in the Machine Learning section of Frontiers in Artificial Intelligence, is an open-access publication, and the final formatted version is slated to appear soon. For researchers drowning in literature synthesis, the tool offers a concrete proposition: hand it a question and an outline, supply the evidence, and receive a traceable first draft that still belongs, decisively, to the human author.

Subject of Research: An AI-powered retrieval-augmented document writing tool that generates structured, cited drafts from user queries and source literature

Article Title: Discourse2Draft: an AI-powered structured document writing tool

Article References: Talukder, A., Combs, P. K., Schmitt, C. P., Reif, D. M., & Auerbach, S. S. (2026). Discourse2Draft: an AI-powered structured document writing tool. Frontiers in Artificial Intelligence, 9, 1945655. https://doi.org/10.3389/frai.2026.1945655

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1945655

Keywords: Discourse2Draft, artificial intelligence, retrieval-augmented generation, large language models, document writing, PubMed, citations, scientific writing, NIEHS, machine learning, LLM agents, research workflow