- 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