stanford-oval/storm

Built on the DSPy framework, STORM executes a two-stage pipeline: an Internet-grounded pre-writing phase that generates outlines through multi-perspective question asking and simulated expert conversations, followed by a writing stage that synthesizes full articles with citations. Co-STORM extends this with human-AI collaborative discourse, introducing a moderator agent and dynamic mind-map visualization to enable interactive knowledge curation. The system supports pluggable retrieval backends (You.com, Bing, vector stores, Tavily, etc.) and language models via LiteLLM, allowing customization across different cost-quality tradeoffs.

28,001 stars. No commits in the last 6 months.