Report Abuse

Basic Information

STORM is a research-grade LLM system for automated knowledge curation that conducts Internet-based research and writes Wikipedia-style articles with citations. The repository implements two related engines: STORM, which breaks article generation into a pre-writing research stage and a writing stage, and Co-STORM, which adds a collaborative discourse protocol enabling multiple LLM agents and human participants to cooperatively explore a topic. The code provides ready-to-run Runner classes (STORMWikiRunner and CoStormRunner), example scripts, command-line options, and instructions to install the Python package knowledge-storm. The project also releases datasets used in evaluations and includes a live research preview that tens of thousands of people have tried. The repo is intended both to reproduce the STORM/Co-STORM papers and to serve as a customizable platform for experimenting with retrieval, multi-LM setups, and human-in-the-loop knowledge curation.

Links

Categorization

App Details

Features
The codebase implements a modular pipeline with four pluggable modules: knowledge curation, outline generation, article generation, and article polishing. The STORM pipeline uses perspective-guided question asking and simulated conversations to elicit deep, broad research questions grounded in web retrieval. Co-STORM adds a collaborative discourse protocol with distinct agent roles (LLM experts, a moderator, and a human user) plus dynamic mind map construction to organize concepts. The package exposes Runner classes and configurable LM and retriever components, supports many retrieval backends (YouRM, BingSearch, VectorRM, SerperRM, BraveRM, SearXNG, DuckDuckGoSearchRM, TavilySearchRM, GoogleSearch, AzureAISearch) and integrates with litellm-supported language and embedding models. The repo includes examples, scripts, datasets (FreshWiki, WildSeek), and instructions for replicating paper experiments.
Use Cases
For researchers and advanced users interested in automated research, STORM automates the pre-writing research workflow and synthesizes outlines and full-length, citation-backed articles to speed drafting. Experienced Wikipedia editors reported the system is helpful in a pre-writing stage where it surfaces coverage, diverse perspectives, and candidate citations. Co-STORM supports human-AI collaborative exploration by managing turn-taking, generating thought-provoking questions, and maintaining a shared mind map to reduce cognitive load during long discussions. The modular interfaces let users swap retrieval or LM components, ground on custom corpora, and reproduce published experiments. Example scripts and a pip-installable package make it straightforward to run, customize, and evaluate topic curation workflows at scale.

Please fill the required fields*