bsearch

Local-first semantic search for macOS. bsearch indexes the files on your Mac — documents, PDFs, and (eventually) emails and images — and lets you and your AI tools search them semantically, entirely locally.

• Hybrid search — semantic (embeddings) + keyword (BM25), fused.
• Agent-friendly — pyramid summaries and an MCP server so AI agents can survey results cheaply and drill down only where it matters.
• Local-first — bring your own OpenAI-compatible inference server (LM Studio, Ollama, …). Content stays on machines you control. No telemetry.
• Laptop-aware — background indexing that respects battery and stays out of your way; search stays snappy.

Built with Go, SQLite (+ sqlite-vec + FTS5), and boring technology on purpose.

Status

Early. Markdown indexing and semantic search work end to end, served by a daemon that indexes in the background over a unix socket:

make install-agent                   # start the daemon at login, and keep it running
bsearch search "heat pump quote"     # search is a client of the daemon

There is no indexing command — the daemon finds new and changed files by itself, and re-embeds the corpus if you change the embedding model. Keyword/hybrid search, summaries, MCP, and filesystem-event-driven freshness are not built yet. Running the daemon is documented in docs/daemon.md; the full design — decisions, rationale, and rejected alternatives — is in DESIGN.md. Work is tracked in GitHub milestones M1–M7.

Development

Requires Go (see go.mod) and a C toolchain — bsearch is cgo: SQLite and sqlite-vec are linked statically, so builds are native, never cross-compiled.

Dev tools are pinned in mise.toml (mise):

make tools    # mise install — fetch the pinned dev tools
make all      # lint + test + build
make test     # unit tests
make fmt      # gofumpt + goimports

CI is documented in docs/ci.md; the changelog policy every PR follows is in docs/changelog.md.

License

MIT