# STORM — AI Research Report Generator by Stanford

> Stanford's LLM-powered system that researches any topic and writes a full Wikipedia-style article with citations. Simulates multi-perspective expert conversations.

## Install

Copy the content below into your project:

# STORM — AI Research Report Generator by Stanford

## Quick Use

```bash
pip install knowledge-storm

# Set API keys
export OPENAI_API_KEY=sk-...
export YDC_API_KEY=...  # You.com search API
```

```python
from knowledge_storm import STORMWikiRunnerArguments, STORMWikiRunner, STORMWikiLMConfigs
from knowledge_storm.lm import OpenAIModel

# Configure LLMs
lm_configs = STORMWikiLMConfigs()
lm_configs.set_conv_simulator_lm(OpenAIModel(model="gpt-4o-mini"))
lm_configs.set_article_gen_lm(OpenAIModel(model="gpt-4o"))

# Configure and run
args = STORMWikiRunnerArguments(output_dir="./results")
runner = STORMWikiRunner(args, lm_configs)

# Generate a research report
runner.run(
    topic="Impact of large language models on scientific research",
    do_research=True,
    do_generate_outline=True,
    do_generate_article=True,
    do_polish_article=True,
)
runner.post_run()  # Saves article, outline, references
```

## Intro

STORM (Synthesis of Topic Outlines through Retrieval and Multi-perspective question asking) is a research system by Stanford OVAL that writes Wikipedia-quality articles on any topic. It doesn't just summarize — it simulates conversations between multiple AI "experts" with different perspectives, then synthesizes their insights into a structured, cited article.

The process mirrors how human researchers work: discover perspectives, ask deep questions from each angle, gather sources, outline, write, and polish. Output includes a full article with inline citations, a reference list, and a conversation log showing the research process.

With 28,000+ stars, STORM is the leading open-source tool for AI-powered research synthesis.

## Details

### How STORM Works

```
1. PERSPECTIVE DISCOVERY
   → Given topic "quantum computing applications"
   → Identifies 5-8 expert perspectives:
     - Physicist, Computer Scientist, Industry Analyst,
       Ethics Researcher, Quantum Engineer...

2. MULTI-PERSPECTIVE RESEARCH
   → Each "expert" asks questions from their angle
   → Questions are answered using web search (You.com, Bing)
   → Simulated conversations deepen understanding

3. OUTLINE GENERATION
   → Synthesizes all research into a structured outline
   → Identifies key sections, subsections, themes

4. ARTICLE WRITING
   → Writes each section using gathered evidence
   → Adds inline citations [1], [2], [3]
   → Maintains coherent narrative across sections

5. POLISH
   → Removes redundancy, improves flow
   → Verifies citation accuracy
   → Outputs final article + references
```

### Output Structure

- `storm_gen_article.txt` — Full article (2,000-5,000 words)
- `storm_gen_outline.txt` — Structured outline
- `url_to_info.json` — All source URLs with extracted content
- `conversation_log.json` — Full simulated expert conversations

### Supported LLM Backends

- OpenAI (GPT-4o, GPT-4o-mini)
- Anthropic (Claude)
- Ollama (local models)
- Any LiteLLM-compatible provider

### Co-STORM: Collaborative Mode

Co-STORM lets you participate in the research process interactively. The AI researches while you steer the direction, ask follow-up questions, and refine the scope in real-time.

## Frequently Asked Questions

**Q: How long does a full research report take?**
A: 3-10 minutes depending on topic complexity and LLM speed. Most time is spent on web search and multi-perspective conversations.

**Q: How accurate are the citations?**
A: Citations link to real web sources. STORM extracts and verifies information from each source. However, always verify critical claims — it's AI-assisted research, not a replacement for peer review.

**Q: Can I use it for academic papers?**
A: It's excellent for literature reviews, background research, and first drafts. The output needs human review and editing for publication-quality work.

**Q: Does it work offline?**
A: Partially. With Ollama for LLM and cached sources, the generation works offline. But the research phase needs internet for web search.

## Works With

- OpenAI, Anthropic, Ollama, any LiteLLM provider
- You.com Search API, Bing Search API
- Python 3.11+
- Any topic — technical, historical, scientific, current events

## Source & Thanks

- **GitHub**: [stanford-oval/storm](https://github.com/stanford-oval/storm) — 28,000+ stars, MIT License
- **Paper**: [Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models](https://arxiv.org/abs/2402.14207)
- By Stanford OVAL (Open Virtual Assistant Lab)

---
<!-- ZH -->
# STORM — 斯坦福 AI 研究报告生成器

## 快速使用
```bash
pip install knowledge-storm
```
```python
from knowledge_storm import STORMWikiRunner
runner = STORMWikiRunner(args, lm_configs)
runner.run(topic="你的研究主题", do_research=True, do_generate_article=True)
```

## 介绍
斯坦福 OVAL 实验室的 AI 研究系统。给定任何主题，模拟多视角专家对话进行深度研究，然后生成带引用的维基百科级文章。28,000+ stars，AI 驱动研究综合的标杆项目。

## 来源与感谢
- **GitHub**: [stanford-oval/storm](https://github.com/stanford-oval/storm) — 28,000+ stars, MIT 许可证
- **论文**: arxiv.org/abs/2402.14207


---
Source: https://tokrepo.com/en/workflows/f09e9348-2bdc-45aa-9907-cfb30290f69b
Author: Skill Factory