The PDF chatbot that cited page numbers, then moved on

What it does ResearchGPT is a thin FastAPI application that ingests PDFs—either uploaded or fetched from a URL—splits the text into chunks, stores embeddings in Redis, and answers natural-language questions using the OpenAI API. It returns the generated response alongside the specific source excerpt and page number it drew from, so you can verify the LLM isn’t hallucinating the bibliography.

The interesting bit This is a textbook RAG pipeline built from off-the-shelf parts: chunk PDF text, embed it into Redis, retrieve relevant snippets, and prompt OpenAI with context. The author has since stopped maintenance in favor of a free successor service, so the repo now serves as a well-starred snapshot of how a simple retrieval stack fits together.

Key highlights

• Returns answers with specific page citations and source text snippets
• Uses Redis as a local embedding cache, keeping retrieval in-house
• Supports both URL-linked and uploaded PDFs
• Built on FastAPI with roughly 3,500 GitHub stars
• MIT licensed, though no longer actively maintained

Caveats

• Requires an OpenAI API key and a running Redis instance; it is not a standalone or offline tool
• The repository is no longer maintained, and the author now points users to a successor service that is free but not yet open-sourced
• The demo video showcases features exclusive to the live demo at dara.chat, not necessarily the open-source codebase

Verdict Worth a look if you want a minimal reference implementation of PDF Q&A with citations, or if you’re studying how a simple LLM retrieval stack is glued together. Skip it if you need active maintenance, local models, or a production-ready backend.