ResearchOS - AI Research Assistant

ResearchOS is an AI-powered research assistant that turns the overwhelming stream of scientific papers into an interactive, citation-aware knowledge base. It runs semantic search, citation-grounded chat, side-by-side paper comparison, PDF question-answering, and recommendations across arXiv and PubMed. What sets it apart is where the work happens: every embedding — for search, similarity, RAG retrieval, and recommendations — is computed on an AMD Radeon RDNA3 GPU (gfx1100) using a ROCm build of PyTorch and sentence-transformers. Benchmarked against CPU, throughput scales from roughly 85x at small batches to 458x at batch size 500, and the fact that the speedup grows with batch size demonstrates genuine GPU parallelism rather than a fixed offset. This acceleration is what makes the semantic layer fast enough to feel interactive. Users search by meaning instead of keywords, then ask questions whose answers are retrieved from the full PDF text — not just abstracts — and returned with inline citations, so nothing is hallucinated. They can compare two to five papers' methods and contributions, upload their own PDFs to question them, and receive recommendations built from a semantic centroid of papers they already value. The system is a cleanly layered, dependency-injected FastAPI backend with ChromaDB for persistent vector storage and Fireworks AI for LLM inference, paired with a Next.js frontend deployed on Vercel. The AMD ROCm notebook backend is exposed to the web via a Cloudflare Tunnel, and the project is fully open-source with documented GPU benchmark evidence.