1 Local Assistant
Joël Jon Moix edited this page 2026-10-27 11:27:07 +01:00

Local assistant (optional)

The assistant answers questions about one campaign's notes — "where did we last see the cartographer?" — with citations you can reopen. It runs entirely on your computer.

What it is and is not

  • Off by default, enabled per campaign by you.
  • Runs on your graphics card (NVIDIA) through Ollama, with cloud features verified off. If no suitable GPU is available, it refuses before sending any note to the model. There is no cloud fallback.
  • Uses only the notes of the campaign you ask about and the wiki entries you approved.
  • Never changes notes, relationships, the sheet or inventory.
  • Models: Qwen3.8-27B (primary) and Gemma 4 12B (secondary). They are general multilingual models, not trained on D&D. A valid citation does not make an answer true — check the source.

Setting it up

  1. App settings → Local AI: components installs Ollama and the chosen model from their official sources, with integrity checks. Downloads start only when you press the button.
  2. In the campaign, open Local assistant, check the runtime and activate the model.

Campaign retraining (advanced)

With Gemma, Campaign settings → Local AI → Retrain on new data prepares a small campaign-specific adapter on your GPU. It is activated only when you choose so. Update knowledge refreshes the search index without retraining. Adapters are included in backups; model weights are not.