Local AI with Ollama

Kept’s AI features (filling things in from photos and receipts, the assistant, semantic search) work with whichever provider a location’s owner connects in Settings → AI. Ollama runs the model on your own hardware, so nothing leaves it; Kept talks to it as an OpenAI-compatible provider.

Needs more than the floor

Ollama on the 2 GB floor hasn’t been measured yet. Until it is, plan for a machine with more memory than Kept alone needs, and expect a small model to be slower and less accurate than a hosted one at reading receipts.

Setting it up

1. Start Kept with the ollama profile, then pull a model into it with Ollama’s own command:

   docker compose --profile ollama up -d
   docker compose exec ollama ollama pull <model>

   The models live in the ollama volume. Nothing is published on the host: only Kept, on Compose’s network, reaches Ollama. The image is about 3.8 GB (amd64) or 2.8 GB (arm64) before any model.

2. In Settings → AI, add an OpenAI-compatible provider with the base URL http://ollama:11434/v1 and the model you pulled. Ollama ignores the API key, so any value does.