Hi HN! Minna is a new tool geared towards people who have a knowledge base so big they forget where things are.- Chat answers are constrained to retrieved results. The model has to search your knowledge base first and cite the document it pulled from. It won t answer from outside your knowledge base which means you can trust the LLM instead of having to be weary about what it outputs. - Hybrid retrieval — BM25 over an inverted index for exact and keyword matches, plus a vector index for semantic matches, results are fused using Reciprocal Ranked Fusion. - Embeddings and inference both run on-device. Search uses bge-small-en and run on the neural engine. Chat uses Qwen3 4B running on MLX. - Search latency is ~12 ms across 250 documents split into 30,000 chunks on Apple Silicon.What’s not done; OCR scanning on PDFs is limited, the app is currently Apple Silicon only, and there are some rough edges in the UI.It’s a TestFlight beta (apologies, but it is the cleanest distribution path right now): https://testflight.apple.com/join/6ZxKb2mm