rahulgit24/research-paper-rag-system

This is a backend API for asking questions about research papers using AI. You upload PDFs, and the system stores them in a way that allows you to ask natural-language questions and get back answers drawn from the actual text of those papers. It is described as production-ready, meaning it is built with care around real problems that come up when more than one person uses a system like this. The interesting part is the multi-step process used to find the most relevant text before answering. A basic approach would be to convert the question into a number and find paper sections with similar numbers, a technique called vector search. This system adds two more filtering steps on top of that. First it uses keyword matching to complement the vector results, then it uses a more expensive ranking model to compare each candidate passage against the question and pick the best few. The result is that only the most genuinely relevant passages reach the AI, which produces better answers than vector search alone. All of this runs on a regular CPU with no specialized graphics hardware needed. The system also handles multi-user sharing thoughtfully. When two people upload the same PDF, the document is only analyzed and embedded once. Both users get access to the same underlying data, but if one of them later deletes the document, the other still keeps their copy. The deletion only removes that user's access, the shared data is only fully deleted when no one needs it anymore. Follow-up questions in a conversation are handled with a query rewriting step. If you ask "what does it say about the training data?" after a question about a specific paper, the system rewrites your vague follow-up into a self-contained question before searching, so references like "it" or "this" resolve correctly. The API requires PostgreSQL for metadata and document tracking, Qdrant as the vector database, and a Groq API key for the language model that generates answers.