Ollama Setup (Local/Offline)

For developers desiring 100% privacy, zero network latency fees, and offline development capabilities, AI Scene Organizer integrates with Ollama. Ollama allows you to host state-of-the-art open-weights models directly on your hardware.

1. Setting Up Ollama

Download and install Ollama from Ollama.com.
Run the Ollama app (you should see the Ollama icon in your taskbar/tray).
Open your terminal (Command Prompt, PowerShell, or macOS Terminal) and download a model capable of structure and reasoning:
```
ollama pull llama3
```
(Alternatively, you can pull other models like ollama pull deepseek-r1 or ollama pull mistral)
Verify the server is running by visiting http://localhost:11434 in your browser. You should see “Ollama is running”.

2. Configuring Unity Settings

In Unity, navigate to Edit ➔ Project Settings... ➔ AI Scene Organizer.
Set the Active Provider dropdown to Ollama.
Set the Ollama API Endpoint (defaults to http://localhost:11434).
Set the Ollama Custom Model ID to match the model you pulled (e.g., llama3, deepseek-r1, or mistral).
Click Apply & Save Settings.

3. Recommended Local Models

Because scene reorganization relies on JSON formatting, we recommend models with strong command compliance:

llama3 (or llama3.1): Highly stable instruction adherence, outstanding structural JSON output. (Recommended default)
deepseek-r1: Excellent reasoning abilities, though thinking tags (<think>...</think>) are filtered by the Unity client to extract clean commands.
mistral: Good lightweight option for smaller scene sorting workloads.

4. Troubleshooting Local Ollama

Error Connection Refused: Make sure Ollama is actually running in your system tray. Try running ollama run llama3 in your terminal to force start the service.
Slow execution times: Running models locally depends entirely on your GPU/CPU capabilities. If scene organization hangs for more than 5 seconds, consider using a smaller quantized model or switching to a cloud provider.
JSON Formatting Errors: Small 1B-3B parameters models may occasionally output malformed JSON. Ensure you are using at least a 7B/8B model (like llama3) for reliable structural hierarchy edits.