Papyrus

Papyrus lets you upload research papers as PDFs and ask questions about them. Answers stream back from Claude with a citation to the paper and page each claim came from. I built it in April 2026.

How a question gets answered

1. Upload. The backend pulls the text out of each PDF with PyMuPDF, splits it into chunks of about 1,000 characters that remember which section they came from, embeds them with a sentence-transformer model (all-MiniLM-L6-v2), and stores them in Qdrant.
2. Retrieve. A question is embedded the same way, and of the 10 nearest chunks, the best 5 go forward.
3. Answer. Claude gets those chunks and a system prompt that requires a citation like [Paper Title, p. 4] for every claim. Its answer streams to the browser over server-sent events, along with the sources it used.

Besides plain questions, it can compare papers side by side, with a few chunks from each, or write a literature review across the library. Claude also pulls the references out of each paper, and a citation graph draws each paper and the papers it cites.

Stack

• Backend: FastAPI, Qdrant, SQLite through SQLAlchemy, and the Anthropic API, with 30 tests on the chunking and PDF processing.
• Frontend: Next.js, React and Tailwind, with pages for uploading, the library, chat and the citation graph.
• Docker Compose runs Qdrant, the backend and the frontend together.