Config

Copy config.example.json to config.json (gitignored).

Env

config.json

Useful knobs (see config.example.json):

• watch.createMissingRoots: create empty watch roots instead of failing doctor
• whisper.promptTerms: names and terms Whisper should prefer
• whisper.computeType: float16, or int8_float16 if VRAM is tight
• audio.preprocess: ffmpeg denoise
• queues.*.concurrency: keep transcription at 1 on a single GPU
• http.tailscale: bind the inbox to this machine's Tailscale address and allow those origins for the UI
• ollanet.required: treat missing cleanup host/model as a failure (default false; raw transcripts still work)
• journal.search: hybrid (default), semantic, or lex. lex matches words only and never embeds note text.
• journal.embedModel: embedding model name. Empty picks the first embed-class model on an allowed host.
• journal.embedHost: where search indexing may send note text (up to 8,000 characters per note) to embed it.

- local (default): this computer only. - machine: the ollanet.machine host, matched the same way cleanup matches it. If ollanet.machine is empty, nothing is embedded. - a host name: that one ollanet host (or alias, or address). - any: this computer first, then any host ollanet discovers from its config, OLLANET_HOSTS, or Tailscale.

If no allowed host has an embedding model, search falls back to words only. pnpm run doctor and the inbox Tools panel show a journal-search check that says so and why.

Where text goes

Audio and transcript files stay on this computer. Cleanup sends transcript text only to ollanet.machine. Search indexing embeds note text on this computer unless you set journal.embedHost to another host.