This repository contains the code base built by the AreYouRedis team to create the "Darwinian Paper Explorer" app.
The hackathon - organised by the MLOps Community, in collaboration with Redis and Saturn Cloud - focused on Vector search engineering, on the arXiv dataset.
For more information on it, please visit the hackathon's welcome page
The submission is a streamlit app, hosted on Saturn Cloud, and accessible by clicking on this link.
The app's vision is to combine vector search and co-citations graph structures in order to:
-
Help the user find arXiv articles linked to a subject he's interested in
-
Provide the user with a view of a topic's trend, based on the number of related publications throughout time, along with a trend prediction for the next two years
-
Illustrate how the subject of interest has evolved throughout history, with an arc diagram highlighting the founding / most influential papers and co-citation relationships between papers
-
Recommend a curated reading list, based on a combination of vector similarity score and paper node degree in the citations graph
It takes as input a user's query - a sentence describing what type of arXiv papers a person is interested in searching - along with a set of filters on the year, categories, and number of similar papers to be retrieved.
Using RedisSearch capabilities, a set of similar articles, linked to the query, is obtained.
Four output sections are then presented:
-
Topic trend & future projection
-
Topic evolution
- Reading list recommender - based on the papers' similarity scores and node degree in the citation graph
- Papers overview
First, clone the repository:
git clone https://github.com/artefactory/AreYouRedis.git
Create a virtual environment at the root of your local repository:
python3 -m virtualenv .venv
source .venv/bin/activate
Install dependencies:
pip install -r requirements.txt
Finally, add the database's password as environment variable:
export REDIS_PASSWORD = '{password}'
To launch the app locally, run the following command:
streamlit run app/app.py
The page below should open in your web browser:
Going from the raw arXiv dataset to the data uploaded in redis incured several steps:
-
The preprocessing operations, as long as utils function to handle the datasets, can be found in the vector.py script.
-
Functions related to data loading, index creation & similarity search are in redis_db.py.
-
Functions used to gather the citations (see next section) are written in scholar_citations.py.
Those different modules are called in the custom-single-gpu-arxiv-embeddings.ipynb notebook, where datasets are gathered, cleaned, merged and loaded into Redis.
To better understand the origin & links between different articles, we chose to gather their citations - i.e. all other arXiv papers cited in one article.
For this, we used Semantic Scholar, an online research tool focused on scientific literature. We worked specifically with the python API, to query the citations.
Finally, the code used to generate the different graphs (topic evolution, prediction of the trend & arc graph) can be found in the topic evolution script.
.
├── LICENSE
├── Makefile
├── README.md
├── app
│ ├── app.py
│ ├── config_files
│ │ └── config.py
│ ├── features
│ │ └── topic_evolution.py
│ ├── main_page.py
│ ├── style
│ │ ├── Artefact_logo.png
│ │ ├── Artefact_small_logo.jpeg
│ │ ├── Redis_logo.png
│ │ ├── Saturncloud_logo.webp
│ │ └── style.css
│ └── utils
│ ├── display.py
│ ├── graph.py
│ ├── load_css.py
│ └── widgets.py
├── entrypoint.sh
├── notebooks
│ ├── custom-single-gpu-arxiv-embeddings.ipynb
│ ├── multi-gpu-arxiv-embeddings.ipynb
│ └── single-gpu-arxiv-embeddings.ipynb
├── requirements.in
├── requirements.txt
├── setup.py
└── src
├── categories.py
├── config.py
├── models.py
├── redis_db.py
├── scholar_citations.py
└── vectors.py