STORM

STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective question asking) automates the pre-writing and drafting stages of long-form research: it surveys related articles to discover distinct perspectives, simulates grounded conversations between a writer and topic experts, builds an outline, and then writes a full-length article in which every claim carries a citation. The companion Co-STORM mode adds a collaborative discourse protocol with LLM experts, a moderator agent, and a dynamic mind map so humans can steer the research as it happens. Built by Stanford's OVAL lab on the DSPy framework with a modular pipeline, it installs as pip install knowledge-storm, supports any model through LiteLLM, and plugs into retrievers including Bing, You.com, Serper, Brave, Tavily, DuckDuckGo, SearXNG, Azure AI Search, and vector search. Code is MIT licensed with about 30,000 GitHub stars, the research appeared at NAACL 2024 with a Co-STORM follow-up paper, and more than 70,000 people have tried the free research preview at storm.genie.stanford.edu.