evanhu1/talk2arxiv

What it solves

Talk2Arxiv provides a seamless way to read and interact with academic papers from arXiv and bioRxiv. It transforms a standard paper link into an interactive experience where users can chat with an AI that has full context of the paper's content, including equations and figures, without needing to upload files manually.

How it works

The system uses a React-based frontend hosted on Vercel and a Cloudflare Worker API. When a user visits a talk2arxiv.org or talk2biorxiv.org link, the API fetches the paper's HTML or PDF. Instead of using RAG (chunking or vector databases), the project leverages large context windows (up to 800K tokens) to feed the entire paper text into the model's context, utilizing OpenAI's prompt cache for efficiency. It uses GPT-6 Luna via OpenRouter to generate responses.

Who it’s for

Researchers, students, and academics who need to quickly summarize, analyze, and explain complex scientific papers from arXiv and bioRxiv.

Highlights

• Interactive Reading: Select text passages to ask the AI for explanations or specific questions.
• Author Context: Automatically retrieves an index of the authors' other works via OpenAlex to provide broader research context.
• Citation Previews: Allows users to preview cited papers and add them to the chat context alongside the original paper.
• PDF Fallback: Uses EmbedPDF (PDFium in WebAssembly) to render papers that lack an HTML version, while maintaining AI chat capabilities.
• Efficient Caching: Implements answer caching via Cloudflare Cache API to reuse identical questions across users.
• Local History: Saves conversation history in the browser's localStorage for persistence.