Local Agentic Copilot
Cloud copilots send private documents off-device and often answer without showing where facts came from.
A fully local copilot on Ollama (Qwen3) with Researcher, Coder and Reviewer agents, deterministic guardrails and a bounded self-correction loop.
Reproducible benchmarks compare Top-1 and Top-3 retrieval across semantic, BM25, hybrid and reranked setups, with failure analysis driving each iteration. On a 47-question benchmark, dense and BM25 each reached 91.5% Top-1 but failed on different questions; hybrid fusion (RRF) reached 100% (47/47), and a cross-encoder reranker added ~50 ms/query with no further gain.
