STORM

What it is

STORM is a language model system that writes Wikipedia-like articles from scratch, based on Internet search. The Stanford Open Virtual Assistant Lab made it, and the name stands for Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. A pre-writing stage collects references and writes an outline. A writing stage turns the outline and the references into a full-length article with citations. Co-STORM, in the same package, lets you take part: you watch or steer a discussion between AI experts and a moderator. The README says the result is not publication-ready, but experienced Wikipedia editors found it helpful in their pre-writing stage. This listing is Mindket’s setup notes for the Python package, its example scripts and its demo interface.

1. You give STORM a topic
2. Simulated writer and expert conversations gather sources
3. It writes an outline from what was collected
4. It writes the article with citations to the sources
5. It polishes the article and saves the files

Who it suits

People who need a cited first draft on a topic they do not know yet, such as students, researchers, writers and editors. Developers who want to build on a knowledge curation engine can change its modules, because the README says STORM is built in a modular way. Skip it if you want a finished article or no setup. You bring a language model and a source to search, and the maker marks the package Alpha.

Where to use it

STORM is Python code that you run from a terminal. The maker gives no steps per operating system. The README documents no skill, plugin or MCP server, so a coding agent can only run its commands for you.

• The knowledge-storm package from PyPI, for Python 3.10 or newer
• A checkout of the repository, which holds the example scripts and the Streamlit demo interface
• Language models: the README says every model that litellm supports. The example scripts cover OpenAI, Azure OpenAI, Anthropic, Google, DeepSeek and Groq, plus VLLM and Ollama servers
• Search sources: You.com, Bing, Brave, Serper, Tavily, SearXNG, DuckDuckGo, Azure AI Search and Google Search, or your own documents through VectorRM

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with a page for each route from the maker’s README and docs. You install the package, give it your keys and run an example script.

1. Install the package. Its requirements pin one exact version of dspy_ai, so use its own environment.
2. To run the example scripts, clone the repository and install its requirements, as the README does with conda.
3. Put your keys in a secrets.toml file in the root folder of the checkout.
4. Run an example script with a retriever and the four stage flags.

pip install knowledge-storm

For a source install, the README gives these lines:

git clone https://github.com/stanford-oval/storm.git
cd storm
conda create -n storm python=3.11
conda activate storm
pip install -r requirements.txt

How to use it

Run the GPT example script from the checkout’s root folder. This is the README’s first example.

python examples/storm_examples/run_storm_wiki_gpt.py \
    --output-dir $OUTPUT_DIR \
    --retriever bing \
    --do-research \
    --do-generate-outline \
    --do-generate-article \
    --do-polish-article

The script asks for a topic and writes its files to a folder named after the topic. The four do flags switch the stages on. A stage you leave out loads its saved results instead, so you can run the research once and rerun only the writing.

Example

You want a cited overview of a topic you have never read about. You put an OpenAI key, the line OPENAI_API_TYPE="openai" and a Bing Search key in secrets.toml. From the checkout’s root folder, you run the README’s first example:

python examples/storm_examples/run_storm_wiki_gpt.py \
    --output-dir $OUTPUT_DIR \
    --retriever bing \
    --do-research \
    --do-generate-outline \
    --do-generate-article \
    --do-polish-article

You type the topic at the Topic prompt. When the run ends, the topic’s folder holds the outline, the article and the sources it used, as the table below shows.

The file url_to_info.json lists the sources the final article used. Read the article as a draft, and check its claims against them. The README says such articles often need a significant number of edits.

For newcomers and experts

If you are new. STORM does what a careful writer does before writing. It asks questions, reads sources, makes an outline and then writes. The README says asking good questions is the core of automating research, and that directly prompting a model to ask questions does not work well. So STORM first looks at articles on similar topics to find different perspectives, then simulates a conversation between a Wikipedia writer and a topic expert, grounded in Internet sources. You supply two outside parts, a language model and a search source. The README also links the maker’s hosted research preview, if you only want to see what it makes. These notes cover the code only.

For experts. The engine has four modules: knowledge curation, outline generation, article generation and article polishing. Each has an interface in knowledge_storm/interface.py and an implementation under knowledge_storm/storm_wiki/modules/. The runner has five language model slots, and the README suggests a cheaper model for the conversation simulator and a stronger one for the article writer. Its example script uses gpt-3.5-turbo for the first two slots and gpt-4o for the other three.

To ground a run on your documents, VectorRM reads a Qdrant vector store that you build from a CSV, offline in a folder or on a Qdrant server. Co-STORM has its own runner with six model slots, a moderator and a turn policy, and the README points to DiscourseManager as the code to change. The maker also publishes two datasets, FreshWiki and WildSeek, and keeps the code behind the STORM paper on its own branch.

Good to know

STORM’s code runs on your machine, and every run reaches out to the model and search services you set.

• Model and search calls: a run sends your topic, its prompts and its search queries to the model service and the search service you configure. The README and the docs say nothing about usage data, and give no run time or cost figures. With a model on a server of yours, such as Ollama or VLLM, the model calls stay with you, and the search queries still go to the search service
• Your documents: VectorRM searches your own CSV in place of web search. In offline mode the vector store stays in a folder on your disk. In online mode the script loads your documents into the Qdrant server you name, such as a Qdrant cloud server. Its example script still calls OpenAI or Azure OpenAI models to write the article
• Keys: secrets.toml holds your keys as plain text in the checkout. Keep it out of any repository you push
• Old model names: the example scripts name models from 2024, such as claude-3 models, gemini-1.5 models and llama3-70b-8192. Check that your account still offers each one, and edit the script when it does not
• Run flags: the GPT script has no default for --retriever and stops with an error without it. Lower --max-thread-num if you keep getting “Exceed rate limit” errors from the model API
• Maturity: setup.py marks the package Alpha, and the README says such articles often need a significant number of edits
• Videos: they come from third parties, and these notes have not checked their steps. Where they differ, follow the steps in these notes
• Removal: the README gives no uninstall steps. Mindket’s notes say what a setup creates, so you can delete it by hand
• License: MIT. The LICENSE file reads “Copyright (c) 2024 Stanford Open Virtual Assistant Lab”