CoinRAG is a side-by-side demo of what the research paper calls nugget-level retrieval — a finer-grained alternative to traditional chunk-based RAG. Instead of retrieving whole chunks and stuffing them into context, the system first extracts concise text-span nuggets offline, then runs two-stage retrieval: chunk FAISS top-k, followed by nugget re-ranking. The payoff is real context compression, using roughly 40 to 60 percent fewer tokens than full chunks on multi-hop queries — while often answering them better.
The demo makes the trade-off visible. It compares traditional RAG against CoinRAG on the same multi-hop questions, showing both answers side by side alongside tokens used. It intentionally skips the raw KV-cache slicing and position-alignment internals, which need model-level access, and instead faithfully reproduces the algorithmic wins: nugget precision, two-stage retrieval, and compact context. The result is a clean, honest look at where retrieval systems are heading.
Check out the demo here:
