LLM / RAG · 2026
Paper Copilot
A literature-review companion. Entirely offline.
The project
Local literature-review copilot — PDF on the left, chat on the right, every answer carrying (paper, page) citations and refusing when evidence is weak. Runs fully offline: no API keys, no GPU. Evaluated properly on 39 papers and 59 hand-reviewed questions with bootstrap CIs and paired McNemar tests, which showed the retrievers are statistically tied (TF-IDF vs hybrid, p = 1.000) and that LLM query-reformulation actively degraded results at 2x latency — built, measured, and left off by default.
At a glance
0.88 Doc-Hit@5 — hybrid RRF, MRR 0.820 — n=59, bootstrap 95% CI.
Built with
RAG · Retrieval Eval · FastAPI · Python