voxora-config

Single source of truth for every voxora runtime setting: the model cache root plus the Hugging Face token, base URL, and default revision. ASR-specific by design — it models what the voxora speech-recognition stack needs, not LLM or vision settings.

Every setting resolves through one cascade, first non-empty wins: an explicit value on VoxoraConfig (set by the caller, or read from a TOML file with VoxoraConfig::from_file), then the environment, then the built-in default.

Variables honoured: VOXORA_CACHE_DIR, VOXORA_HF_TOKEN, VOXORA_HF_BASE_URL, VOXORA_HF_REVISION, plus the HF_TOKEN and HUGGING_FACE_HUB_TOKEN aliases. Per-setting rules: module docs.

Related: hardware backend selection

Hardware backend selection (CUDA / Metal / Vulkan / CPU) is a compile-time Cargo feature per engine, with a runtime voxora_backend::best_device() probe when the candle feature is enabled. See docs/GPU_SUPPORT.md for the full cross-engine matrix and voxora-backend/README.md for the runtime picker.

Examples

• env_cascade — print the resolved cache root, HF token, HF base URL, and default HF revision with every cascade layer applied. Pass an optional positional TOML path to verify a voxora.toml is read at the expected location. Run with cargo run --example env_cascade -p voxora-config (or cargo run --example env_cascade -p voxora-config -- voxora.toml).