stanford-oval/storm

STORM is a research project from Stanford that uses large language models to write Wikipedia-style articles with citations, starting from nothing but a topic name. You hand it a subject, it searches the internet, gathers references, drafts an outline, and then writes a full article that quotes the sources it found. The README is upfront that the result is not publication-ready and still needs editing, but a live demo at storm.genie.stanford.edu has been tried by more than 70,000 people. It works in two stages. In the pre-writing stage, STORM runs internet searches and builds an outline. The questions it asks are not naive, the system first surveys similar Wikipedia articles to discover different perspectives on the subject, and it simulates a conversation between a writer and a topic expert grounded in the search results to ask sharper follow-ups. In the writing stage, it takes the outline and references and drafts the article with citations. A second variant called Co-STORM adds a human-in-the-loop: AI experts, a moderator that asks thought-provoking questions, and the user share a dynamically updated mind map of what has been learned. You would use it when you need a structured first draft on an unfamiliar topic, a literature review, a briefing, a research starter, and want something more thorough than a single chat answer. The code is Python, installable via pip install knowledge-storm, built on the dspy framework, and supports any LLM and embedding model exposed by litellm plus retrievers including Bing, Google, Serper, Brave, Tavily, You.com, DuckDuckGo, SearXNG, Azure AI Search, and a local vector store.