Description
Writing a survey of an unfamiliar field, the time sink is not the writing — it is the legwork before it: searching a dozen sources, judging which to trust, working out which angles to cover, and matching every citation to a claim. STORM, an open system from Stanford's OVAL lab, automates that whole pre-writing stage — give it a topic and it researches the web, builds an outline, and writes a Wikipedia-like article where every sentence carries its source.
Its cleverest move is how it asks questions. Prompting a model to interrogate itself produces shallow, repetitive questions; STORM instead surveys existing articles on similar topics to discover distinct perspectives, then runs multi-turn conversations between a "Wikipedia writer" and a "topic expert" whose answers are grounded in retrieved web pages. The questions that fall out are deeper and broader, and the citations hold up. The follow-on Co-STORM even lets you join the conversation as a human — watch the agents discuss, inject a turn to steer the focus where you actually care — with a live mind map maintained alongside.
More than seventy thousand people have tried the Stanford preview. The whole thing is built modularly on dspy, so the language model and the retrieval source are both swappable.
Two-stage generation: a pre-writing stage researches the web and drafts an outline, then a writing stage produces the full article with citations, splitting long-form writing into two automatable steps.
Perspective-guided questions: surveying similar articles first yields distinct perspectives that steer the questioning, avoiding the shallow self-interrogation a bare prompt produces.
Simulated-conversation research: a writer and an expert exchange multiple turns with every expert answer grounded in retrieved pages, updating understanding and asking follow-ups as they go.
Co-STORM collaboration: join the agents' discussion to observe or steer it, while the system maintains a dynamic mind map that cuts the mental load of a long discourse.
Swappable models and sources: built on dspy, with language models wired in through litellm and retrieval over Bing, You, or your own vector store.
Its cleverest move is how it asks questions. Prompting a model to interrogate itself produces shallow, repetitive questions; STORM instead surveys existing articles on similar topics to discover distinct perspectives, then runs multi-turn conversations between a "Wikipedia writer" and a "topic expert" whose answers are grounded in retrieved web pages. The questions that fall out are deeper and broader, and the citations hold up. The follow-on Co-STORM even lets you join the conversation as a human — watch the agents discuss, inject a turn to steer the focus where you actually care — with a live mind map maintained alongside.
More than seventy thousand people have tried the Stanford preview. The whole thing is built modularly on dspy, so the language model and the retrieval source are both swappable.
Features
Two-stage generation: a pre-writing stage researches the web and drafts an outline, then a writing stage produces the full article with citations, splitting long-form writing into two automatable steps.
Perspective-guided questions: surveying similar articles first yields distinct perspectives that steer the questioning, avoiding the shallow self-interrogation a bare prompt produces.
Simulated-conversation research: a writer and an expert exchange multiple turns with every expert answer grounded in retrieved pages, updating understanding and asking follow-ups as they go.
Co-STORM collaboration: join the agents' discussion to observe or steer it, while the system maintains a dynamic mind map that cuts the mental load of a long discourse.
Swappable models and sources: built on dspy, with language models wired in through litellm and retrieval over Bing, You, or your own vector store.
