ArXiv Vectors
Last summer, while working on a project, I spent ages searching for relevant papers. At one point, I wished for a magic oracle that I could ask about any paper. Luckily, we have embeddings and vector databases. They turned my wish into reality, helping me create my own oracle.
You can try it yourself here!
Frankly, there are a lot of CS/ML papers on arXiv. Perhaps too much. There are projects such as arXiv sanity which tries to alleviate this issue, but with arrivial of vector databases and high-quality LLM embeddings, I thought it’d be fun to quickly build a search engine for arXiv papers. Suprisingly it only costs $6 to embed all CS/ML abstracts from arXiv using OpenAI embedding service. I’ve used pinecone for vector database, Svelte for building the application, and fly.io for hosting.