Most tools that promise "AI search over your notes" quietly ship those notes to a cloud API. This one takes the opposite bet: keyword ranking, embeddings, and the final LLM re-rank all run on your own machine, so a private corpus never leaves the device. The interesting claim is that good hybrid retrieval no longer needs a server behind it.
What Sets It Apart
- Three-stage retrieval, not one trick. BM25 catches exact terms, vector search (default EmbeddingGemma-300M, swappable for multilingual models like Qwen3-Embedding) catches meaning, and a local LLM re-ranks the survivors — so a vague query like "the meeting where we cut the budget" can still land.
- Genuinely offline. Everything runs through node-llama-cpp with GGUF models on top of SQLite (FTS5 + sqlite-vec): no API keys, no network round-trips, no per-token billing.
- Built for agents, not just humans. Beyond the CLI it exposes an MCP server and emits JSON, CSV, Markdown, and XML, so coding agents and scripts can treat your notes as a queryable tool.
- Content-aware chunking. Prose is split into overlapping windows (~900 tokens, 15% overlap) while code is chunked along its AST, so retrieved fragments stay coherent instead of being cut mid-function.
Who It's For
Great fit if you keep a large, private corpus — markdown notes, meeting transcripts, a codebase — and want agent-ready retrieval without handing any of it to a third party. Look elsewhere if you need hosted multi-user team search or zero local setup: it expects you to run models locally, and indexing a big corpus on-device costs CPU, RAM, and time up front.
