Regenerable local indexes

Each vault's libSQL index is a derived projection of its markdown and Lyt declarations. It is local to the machine, can be rebuilt, and never becomes the authority over the source files. A pod search combines the indexes of vaults available in that pod. Both the indexing and the searching run on your machine; your notes are not sent anywhere to be searched or indexed.

A confidence cascade

lyt search works across the pod by default; use its--vault option to focus on one vault. Search gathers candidates from several local signals at once and blends them best-first: notes the pod has already organized around a topic rank ahead of raw text matches, and full-text retrieval backfills everything else.

Matching by meaning

Matching by meaning is on by default, but it needs a small language model that runs on your own computer. Downloading that model once is a separate choice: Lyt asks before fetching it, and running a search never triggers the download. Until it is installed, search works from keywords alone.

Once it is installed, meaning-based matches join the same ranked list, so a note that uses different words for the same idea can surface next to the literal matches. It widens a keyword result rather than replacing one: a query still has to match some text before meaning is consulted, and meaning-only matches fill result slots that keyword matching left empty — a query that already fills the page with literal hits may show none. lyt search --no-semantic turns it off for a single query.

That one-time model download is the only outward request search makes. It fetches the model from a public bucket and carries none of your notes with it.

Search for people and agents

The same local records can be searched through the CLI or the MCP server. An agent works with the pod it has been given; calling an external model remains an explicit adapter choice outside the index itself.