STORM

STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) is a knowledge-curation system from Stanford OVAL that takes a single topic prompt and produces a full long-form report with inline citations. It works in two phases: a pre-writing stage where it searches the web, simulates multi-perspective expert conversations, and builds an outline; and a writing stage where it drafts a structured article grounded in the gathered sources. A free hosted preview lives at storm.genie.stanford.edu, and a Co-STORM variant adds a human-in-the-loop mind map for collaborative research sessions.

It is best understood as a research-grade alternative to consumer 'deep research' features, aimed at writers, analysts, students, and Wikipedia-style contributors who care about traceable sources more than slick prose. The hosted demo is free with terms-of-service signup; the real product is the MIT-licensed Python package (`pip install knowledge-storm`) built on DSPy, which lets you swap in your own LLM provider via LiteLLM and your own retriever (Bing, You.com, Tavily, DuckDuckGo, or a local VectorRM over your documents). Running it seriously means bringing your own API keys and paying for search and model calls.

The trade-off is honest: outputs are encyclopedic and dense rather than narrative, hallucinations and citation drift still occur, and the hosted preview can be slow or queue-limited under load. But as a self-hostable, model-agnostic agent for grounded long-form synthesis, there is little else with comparable academic provenance.