STORM is a cutting-edge AI system designed to automate the process of research and writing, particularly for generating long-form articles with citations. Developed by Stanford's Open Virtual Assistant Lab (OVAL), STORM breaks down the complex task of article generation into two primary stages: the pre-writing stage and the writing stage.
In the pre-writing stage, STORM conducts Internet-based research to collect references and generates a structured outline. The system employs advanced strategies like Perspective-Guided Question Asking and Simulated Conversation to ensure depth and breadth in the research process. By surveying existing articles and simulating expert conversations, STORM refines its understanding of the topic and asks more insightful follow-up questions.
In the writing stage, STORM uses the generated outline and collected references to produce a full-length article with proper citations. The system is modular and customizable, allowing users to integrate different language models (via litellm) and retrieval modules (such as YouRM, BingSearch, and VectorRM).
STORM also introduces Co-STORM, a collaborative extension that enables human-AI teamwork in knowledge curation. Co-STORM features a dynamic mind map to organize information hierarchically, reducing cognitive load during in-depth discussions. It supports a turn-based discourse protocol involving LLM experts, a moderator, and human users, fostering a shared conceptual space.
The project is highly modular, built using dspy, and supports various language and embedding models. It has been used by over 70,000 people in its research preview phase, demonstrating its potential to assist Wikipedia editors and other knowledge workers. STORM is open-source, with active development focused on enhancing human-in-the-loop functionalities and information abstraction capabilities.
To get started, install the package via pip install knowledge-storm or clone the GitHub repository for direct customization. The system is particularly useful for pre-writing stages, though it may require human edits for publication-ready articles.