CiteCheck turns a library of research PDFs into a focused workspace for asking questions, comparing findings, and drafting peer-review style critiques. Answers are grounded in retrieved source passages, with inline citations that make the evidence easy to inspect. The product was built around a simple premise: research assistance should be useful without asking people to trust unsupported output.
CiteCheck runs a second verification pass after the answer is written. It splits the response into atomic claims and checks each one against the retrieved passages, labeling it supported, partially supported, or unsupported. This keeps the interface centered on the evidence behind a response, giving researchers a clear view of what the available source material actually supports.
The FastAPI application combines local ChromaDB embeddings, PDF parsing, OpenRouter-compatible language models, SQLite accounts, and a trained scikit-learn classifier that identifies a passage's rhetorical role. Beyond grounded chat, the workspace supports cross-paper comparison, a multi-agent peer-review panel, translation, per-account data isolation, and rate limits designed for a practical free-tier deployment.