Storm
title: STORM radar_quadrant: Tools radar_ring: Assess radar_position: inner created: 2026-05-22 last_updated: 2026-05-22 related: ["open-parse", "DataContractsScraping"]
STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) is a research system from Stanford OVAL that takes a topic and generates a full-length Wikipedia-style report with citations. It combines internet retrieval with structured multi-perspective questioning to produce coherent long-form articles rather than passive summaries.
STORM operates in two stages:
Pre-writing: The system surveys existing articles on similar topics to identify distinct perspectives, then simulates dialogues between a writer and domain experts to generate follow-up questions that surface non-obvious angles. Internet search (You.com, Bing, or custom vector stores) retrieves sources keyed to these questions. An outline is assembled from the retrieved material.
Writing: The outline and citations are used to generate the full article section by section, with inline citations grounding each claim.
The multi-perspective questioning step distinguishes STORM from naive retrieve-and-summarize approaches: it actively constructs questions a human expert would ask, rather than relying on the original topic query alone.
A collaborative variant (Co-STORM) adds a human-in-the-loop mode with a discourse protocol involving multiple LLM experts, a moderator, and the user. The human can steer research direction, add questions, or redirect focus mid-session.
pip install knowledge-stormSupports multiple LLMs via LiteLLM integration (OpenAI, Anthropic, and others). Search backends are configurable.
Wikipedia editors who tested STORM report the pre-writing and outline stages are useful, but final output requires substantial revision before publication. The system is best understood as a structured research accelerator, not a finished-article generator.
STORM sits at Tools → Assess inner. The outline-first, multi-perspective questioning pattern is a meaningful step beyond naive RAG summarization and applies to technical research, due diligence, and knowledge synthesis tasks beyond Wikipedia. The Co-STORM variant makes it relevant for interactive research workflows. The trial gate is a completed research run on a real topic with verified citation quality and outline coherence before accepting generated prose.
