tinghao-ai/storm

An LLM-powered knowledge curation system that researches a topic and generates a full-length report with citations.

STORM: Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking

This repository contains the code for our NAACL 2024 paper Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models by Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Peter Xu, Omar Khattab, and Monica S. Lam.

Overview (Try STORM now!)

STORM is a LLM system that writes Wikipedia-like articles from scratch based on Internet search.

While the system cannot produce publication-ready articles that often require a significant number of edits, experienced Wikipedia editors have found it helpful in their pre-writing stage.

Try out our live demo to see how STORM can help your knowledge exploration journey and please provide feedback to help us improve the system 🙏!

Research Before Writing

STORM breaks down generating long articles with citations into two steps:

1. Pre-writing stage: The system conducts Internet-based research to collect references and generates an outline.
2. Writing stage: The system uses the outline and references to generate the full-length article with citations.

STORM identifies the core of automating the research process as automatically coming up with good questions to ask. Directly prompting the language model to ask questions does not work well. To improve the depth and breadth of the questions, STORM adopts two strategies:

1. Perspective-Guided Question Asking: Given the input topic, STORM discovers different perspectives by surveying existing articles from similar topics and uses them to control the question-asking process.
2. Simulated Conversation: STORM simulates a conversation between a Wikipedia writer and a topic expert grounded in Internet sources to enable the language model to update its understanding of the topic and ask follow-up questions.

Based on the separation of the two stages, STORM is implemented in a highly modular way (see engine.py) using dspy.

Setup

We view STORM as an example of automated knowledge curation. We are working on enhancing our codebase to increase its extensibility. Stay tuned!

Below, we provide a quick start guide to run STORM locally to reproduce our experiments.

1. Install the required packages.

   conda create -n storm python=3.11
   conda activate storm
   pip install -r requirements.txt

2. Set up OpenAI API key and You.com search API key. Create a file secrets.toml under the root directory and add the following content:

   # Set up OpenAI API key.
   OPENAI_API_KEY=<your_openai_api_key>
   # If you are using the API service provided by OpenAI, include the following line:
   OPENAI_API_TYPE="openai"
   # If you are using the API service provided by Microsoft Azure, include the following lines:
   OPENAI_API_TYPE="azure"
   AZURE_API_BASE=<your_azure_api_base_url>
   AZURE_API_VERSION=<your_azure_api_version>
   # Set up You.com search API key.
   YDC_API_KEY=<your_youcom_api_key>

Paper Experiments

The FreshWiki dataset used in our experiments can be found in ./FreshWiki.

Run the following commands under ./src.

Pre-writing Stage

For batch experiment on FreshWiki dataset:

python -m scripts.run_prewriting --input-source file --input-path ../FreshWiki/topic_list.csv  --engine gpt-4 --do-research --max-conv-turn 5 --max-perspective 5

• --engine (choices=[gpt-4, gpt-35-turbo]): the LLM engine used for generating the outline
• --do-research: if True, simulate conversation to research the topic; otherwise, load the results.
• --max-conv-turn: the maximum number of questions for each information-seeking conversation
• --max-perspective: the maximum number of perspectives to be considered, each perspective corresponds to an information-seeking conversation.
  • STORM also uses a general conversation to collect basic information about the topic. So, the maximum number of QA pairs is max_turn * (max_perspective + 1). 💡 Reducing max_turn or max_perspective can speed up the process and reduce the cost but may result in less comprehensive outline.
  • The parameter will not have any effect if --disable-perspective is set (the perspective-driven question asking is disabled).

To run the experiment on a single topic:

python -m scripts.run_prewriting --input-source console --engine gpt-4 --max-conv-turn 5 --max-perspective 5 --do-research

• The script will ask you to enter the Topic and the Ground truth url that will be excluded. If you do not have any url to exclude, leave that field empty.

The generated outline will be saved in {output_dir}/{topic}/storm_gen_outline.txt and the collected references will be saved in {output_dir}/{topic}/raw_search_results.json.

Writing Stage

For batch experiment on FreshWiki dataset:

python -m scripts.run_writing --input-source file --input-path ../FreshWiki/topic_list.csv --engine gpt-4 --do-polish-article --remove-duplicate

• --do-polish-article: if True, polish the article by adding a summarization section and removing duplicate content if --remove-duplicate is set True.

To run the experiment on a single topic:

python -m scripts.run_writing --input-source console --engine gpt-4 --do-polish-article --remove-duplicate

• The script will ask you to enter the Topic. Please enter the same topic as the one used in the pre-writing stage.

The generated article will be saved in {output_dir}/{topic}/storm_gen_article.txt and the references corresponding to citation index will be saved in {output_dir}/{topic}/url_to_info.json. If --do-polish-article is set, the polished article will be saved in {output_dir}/{topic}/storm_gen_article_polished.txt.

Customize the STORM Configurations

We set up the default LLM configuration in LLMConfigs in src/modules/utils.py. You can use set_conv_simulator_lm(),set_question_asker_lm(), set_outline_gen_lm(), set_article_gen_lm(), set_article_polish_lm() to override the default configuration. These functions take in an instance from dspy.dsp.LM or dspy.dsp.HFModel.

💡 For a good practice,

• choose a cheaper/faster model for conv_simulator_lm which is used to split queries, synthesize answers in the conversation.
• if you need to conduct the actual writing step, choose a more powerful model for article_gen_lm. Based on our experiments, weak models are bad at generating text with citations.

Automatic Evaluation

In our paper, we break down the evaluation into two parts: outline quality and full-length article quality.

Outline Quality

We introduce heading soft recall and heading entity recall to evaluate the outline quality. This makes it easier to prototype methods for pre-writing.

Run the following command under ./eval to compute the metrics on FreshWiki dataset:

python eval_outline_quality.py --input-path ../FreshWiki/topic_list.csv --gt-dir ../FreshWiki --pred-dir ../results --pred-file-name storm_gen_outline.txt --result-output-path ../results/storm_outline_quality.csv

Full-length Article Quality

eval/eval_article_quality.py provides the entry point of evaluating full-length article quality using ROUGE, entity recall, and rubric grading. Run the following command under eval to compute the metrics:

python eval_article_quality.py --input-path ../FreshWiki/topic_list.csv --gt-dir ../FreshWiki --pred-dir ../results --gt-dir ../FreshWiki --output-dir ../results/storm_article_eval_results --pred-file-name storm_gen_article_polished.txt

Use the Metric Yourself

The similarity-based metrics (i.e., ROUGE, entity recall, and heading entity recall) are implemented in eval/metrics.py.

For rubric grading, we use the prometheus-13b-v1.0 introduced in this paper. eval/evaluation_prometheus.py provides the entry point of using the metric.

Contributions

If you have any questions or suggestions, please feel free to open an issue or pull request. We welcome contributions to improve the system and the codebase!

Contact person: Yijia Shao and Yucheng Jiang

Citation

Please cite our paper if you use this code or part of it in your work:

@inproceedings{shao2024assisting,
      title={{Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models}}, 
      author={Yijia Shao and Yucheng Jiang and Theodore A. Kanell and Peter Xu and Omar Khattab and Monica S. Lam},
      year={2024},
      booktitle={Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)}
}