Agent Page
Route: /agent — Source: ui/src/pages/AgentPage.tsx
Chat interface for the in-app assistant. Backed by Gemma 4 (4B) running natively on Apple Silicon via mlx-vlm, with Ollama as a fallback for non-Mac hosts.
Capabilities
| Capability | What it does |
|---|---|
| Insights | Summarize a labeling job, highlight class imbalance, suggest prompts |
| Chat | Answer free-form questions about the current dataset |
| Vision | Analyze a frame or annotation directly (multimodal — pass an image as context) |
| Suggest | Recommend the next action: more frames? new prompt? larger model? |
The same model is wired into the AI Insights drawer on the Review page — the agent panel is just a freer interface to the same backend.
Configuration
| Var | Default | Purpose |
|---|---|---|
AGENT_MODEL_ID | google/gemma-4-e4b-it | Model id for the MLX-VLM path |
OLLAMA_URL | http://localhost:11434 | Fallback Ollama endpoint |
OLLAMA_MODEL | gemma2:4b | Fallback Ollama model |
The page picks the MLX path automatically when mlx-vlm is importable; otherwise it falls back to Ollama. There is no third path — install one or the other.
Privacy
The agent runs locally. Nothing in your dataset, prompts, or frames leaves the host. That's the whole point of running it via MLX or Ollama instead of a hosted API.