Repository navigation
Search your iMessage history by meaning, entirely on your own machine.
uvx msgsearch doctor # try it without installing
pipx install msgsearch # or install the commandApple stores the text of about 86% of messages in an attributedBody blob
rather than in the text column, so a naive search over chat.db finds roughly
one message in seven. msgsearch decodes that blob, then retrieves with BM25 over
conversation windows fused with vector search over three-message passages. Both
models run locally; no message text leaves the machine.
Measured over 9 labelled queries on a 260k-message archive: MRR 0.815,
against 0.657 for keyword search alone and 0.537 for vectors alone. Reranking is
implemented and off by default, because measuring it showed it made results
worse. See ARCHITECTURE.md for the evaluation and CHANGELOG.md for detail.
Requires macOS on Apple silicon and Python 3.10–3.13.