How stanford-oval/storm works
Knowledge-Storm is a modular LLM knowledge-curation system with two products: STORM generates cited Wikipedia-like articles, while Co-STORM supports human-guided research through discourse and a shared mind map. The primary STORM workflow is topic submission, perspective-driven retrieval, outline construction, article writing, polishing, and displayed output. Co-STORM adds warm-start research, moderated turns, grounded answers, and knowledge-base insertion. The graph preserves documented capabilities; several internal module calls are inferred only where the sampled engine or README explicitly describes them.
Main entry points: lm.py Language Models · storm_wiki/engine.py STORM Runner · knowledge_curation.py Knowledge Curation · collaborative_storm/engine.py Co-STORM Runner
Components 23
Orchestration and UI
- Streamlit Demo
frontend/demo_light/pages_util/CreateNewArticle.py - STORM Runner
knowledge_storm/storm_wiki/engine.py - Co-STORM Runner
knowledge_storm/collaborative_storm/engine.py - Article Viewer
frontend/demo_light/demo_util.py
STORM Curation
- Knowledge Curation
knowledge_storm/storm_wiki/modules/knowledge_curation.py - Source Retriever
knowledge_storm/storm_wiki/modules/retriever.py - Information Table
knowledge_storm/storm_wiki/modules/storm_dataclass.py - Outline Generation
knowledge_storm/storm_wiki/modules/outline_generation.py - Article Generation
knowledge_storm/storm_wiki/modules/article_generation.py - Article Polishing
knowledge_storm/storm_wiki/modules/article_polish.py
Co-STORM Collaboration
- Warm Start Research
knowledge_storm/collaborative_storm/modules/warmstart_hierarchical_chat.py - Discourse Manager
knowledge_storm/collaborative_storm/engine.py - Co-STORM Agents
knowledge_storm/collaborative_storm/modules/co_storm_agents.py - Grounded Answers
knowledge_storm/collaborative_storm/modules/grounded_question_answering.py - Information Insertion
knowledge_storm/collaborative_storm/modules/information_insertion_module.py - Shared Knowledge Base
knowledge_storm/dataclass.py
Models and Retrieval
- Retrieval Models
knowledge_storm/rm.py - Language Models
knowledge_storm/lm.py - Embedding Encoder
knowledge_storm/encoder.py
Article Artifacts
- Article Files
frontend/demo_light/demo_util.py
Other components
- Human User
- Web Search
- LLM Provider
Connections 31
- Human User → Streamlit Demo: submits topic
- Streamlit Demo → STORM Runner: runs research
- STORM Runner → Knowledge Curation: starts curation
- Knowledge Curation → Source Retriever: requests sources
- Source Retriever → Retrieval Models: executes queries
- Retrieval Models → Web Search: searches web
- Knowledge Curation → Information Table: collects evidence
- STORM Runner → Outline Generation: generates outline
- Outline Generation → Information Table: uses evidence
- STORM Runner → Article Generation: writes article
- Article Generation → Information Table: uses references
- STORM Runner → Article Polishing: polishes article
- Article Polishing → Article Files: writes result
- Article Viewer → Article Files: reads result
- Streamlit Demo → Article Viewer: shows result
- Co-STORM Runner → Warm Start Research: starts warmup
- Warm Start Research → Grounded Answers: answers questions
- Grounded Answers → Retrieval Models: retrieves grounding
- Warm Start Research → Shared Knowledge Base: initializes map
- Co-STORM Runner → Discourse Manager: manages turns
- Discourse Manager → Co-STORM Agents: selects agent
- Co-STORM Agents → Grounded Answers: requests answers
- Co-STORM Runner → Information Insertion: inserts turn
- Information Insertion → Shared Knowledge Base: updates map
- Shared Knowledge Base → Embedding Encoder: supports similarity
- Knowledge Curation → Language Models: prompts models
- Outline Generation → Language Models: prompts models
- Article Generation → Language Models: prompts models
- Article Polishing → Language Models: prompts models
- Co-STORM Agents → Language Models: prompts models
- Language Models → LLM Provider: calls models