seatedro/arxival

arXival 📚

a research paper answering engine focused on openly accessible ML papers. give it a query, get back ai-generated responses with citations and figures from relevant papers.

what’s under the hood 🔧

data sources

• arxiv + semantic scholar api for paper metadata and pdfs
• paper content processed from pdfs using pymupdf

tech stack

• frontend: next.js 15 + app router, tailwind, shadcn/ui
• backend: fastapi + uvicorn
• vector store: chromadb (running on cloud)
• llm: openai gpt 4o-mini
• embeddings: openai text-embedding-3-large
• storage: cloudflare r2 for extracted figures

running locally 🚀

1. clone the repo:

git clone https://github.com/seatedro/arxival.git
cd arxival

1. set up backend:

cd server
python -m venv .venv
source .venv/bin/activate  # or \`.venv\Scripts\activate\` on windows
pip install -r requirements.txt

1. set up env vars:

# backend (.env in server/)
OPENAI_API_KEY=your_key
PINECONE_API_KEY=your_token
PINECONE_HOST=your_server
R2_ENDPOINT=your_endpoint
R2_ACCESS_KEY_ID=your_key
R2_SECRET_ACCESS_KEY=your_key

1. start the backend:

python run.py

1. set up frontend:

cd ui
npm install

1. set up frontend env:

# frontend (.env in ui/)
NEXT_PUBLIC_BACKEND_URL=http://localhost:8000

1. start the frontend:

npm run dev

1. (optional) ingest some papers:

cd ../server
python cli_batch.py --query "machine learning" --max-papers 50

hit up http://localhost:3000 and you’re good to go! 🎉