ASTRA (ArXiv Sourced Text Recommendation Agent), as the name suggests, is a paper recommendation service based on arXiv. We've collected thousands of papers across various scientific disciplines for recommendation. Simply provide a paper's title, arXiv ID, or a handful of keywords, and Astra will curate a list of relevant documents.
Like many RAG-style applications, Astra relies on quality embeddings of documents and user queries. The core component of our recommendation engine is Embed, a large language model developed by Cohere. Embed was built to handle mixed-modality documents and has demonstrated results in retrieval, making it a perfect choice for Astra. With this approach, recommendation is simply a k-nearest neighbors problem, made expedient with the help of FAISS.
You can find our deployment of Astra here. If you intend on building Astra locally:
git clone https://github.com/colbyendres/Astra/
cd Astra
pip install -r requirements.txt
which pulls in all the necessary dependencies. For generating the embedding datasets, take a look at the notebooks directory. With the database seeded, simply run with flask run.