docs/deploy.md

Deploy

Hugging Face Space (Gradio)

This repo ships a thin Space entrypoint at app.py that imports and runs the runtime package from src/reachy_mini_chat.

Required Space settings

Set Space variables/secrets in the Hugging Face UI.

OpenAI mode:

• REALTIME_PROVIDER=openai
• OPENAI_API_KEY=...
• MODEL_NAME=gpt-realtime

OSS endpoint mode:

• REALTIME_PROVIDER=oss
• REALTIME_BASE_URL=https://your-endpoint.example.com/v1
• REALTIME_API_KEY=...
• MODEL_NAME=mistralai/Voxtral-Mini-4B-Realtime-2602

Optional for both:

• REALTIME_TRANSCRIBE_MODEL (default gpt-4o-mini-transcribe)
• REALTIME_ENABLE_INPUT_TRANSCRIPTION (1 by default)
• REACHY_MINI_ACCESS_KEY (enable auth)

Dependency management

• Canonical dependency list is pyproject.toml ([project].dependencies).
• requirements.txt is generated from it:

python3 tools/sync_requirements.py

Run this before pushing dependency changes to Space.

System packages for Space builds

The default Gradio Space builder ignores this repo's Dockerfile. It installs Python dependencies from requirements.txt and Debian packages from packages.txt.

This repo includes packages.txt with the native libraries required to build PyGObject, which is pulled in by reachy-mini on Linux. Keep that file in sync if runtime dependencies start requiring additional system libraries.