Skip to content

Latest commit

 

History

History
59 lines (45 loc) · 4.82 KB

settings.md

File metadata and controls

59 lines (45 loc) · 4.82 KB

Settings

raea reads each setting from the environment first, then from ~/.raea/env (or the file RAEA_ENV_FILE names): one KEY=value per line, # for comments. Nothing has a hidden default host: a feature that needs a service says which setting is missing.

mkdir -p ~/.raea && touch ~/.raea/env && chmod 600 ~/.raea/env

Network

Setting What
RAEA_OFFLINE=1 raea makes no network requests: a remote URL not already in raea's cache, a model or sample library not yet installed, a composition's own fetch() or WebSocket, Bevy's http asset source and every raea gen step are refused with what wanted out. What's on disk keeps working

Generation (raea gen, optional)

Setting What
FISH_API_KEY Your Fish Audio key, for an audio.speech step on fish:<model> (fish: alone is s2.1-pro); a plan quotes Fish's price for the words first

Local audio models

Two models make sound on your Mac, free, with no account. Each installs itself the first time a composition or a command asks for it (a few minutes, once), or ahead of time with raea setup-audio.

Setting What
RAEA_TOKENSYNTH Where TokenSynth lives (default ~/.cache/raea-tokensynth, about 6 GB): an instrument described in words, one note at a time, for a Tone.js Sampler
RAEA_SAO Where Stable Audio Open Small lives (default ~/.cache/raea-sao, about 2 GB, Core ML): foley and sound effects
RAEA_AUDIO_AUTOSETUP=0 Don't install on first use; a missing model's error names the command (tools/sao/setup.sh on a Mac or Linux, uv run --no-project --python 3.11 python -I tools\sao\setup.py on Windows)
RAEA_SAO_CUDA SAO off a Mac: 1 installs ONNX Runtime with CUDA, 0 the CPU build; unset, an NVIDIA GPU (nvidia-smi) decides
RAEA_TOOLS_CACHE Where raea keeps the tools it fetches when they're not on the PATH (default ~/.cache/raea-tools, each checked against its SHA-256): uv 0.12.24, which every local model's setup runs through; and on Windows FFmpeg 9.0.2 (about 210 MB), fetched the first time a video raea's own decoders can't read is opened (a phone's or a camera's: H.264 with B-frames, HEVC, VP9)

From the command line, raea sao "PROMPT" -o hit.wav and raea tokensynth "TIMBRE" --notes C3,C4 -o notes/ make sounds through the same cache a composition's render uses, and raea sounds searches open sample libraries (pianos, General MIDI instruments, orchestral and percussion samples, each with its licence).

Checks and guides (optional)

Setting What
CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_API_TOKEN The decision model (clef-omni on Workers AI) on your own Cloudflare account, for raea ask, raea judge and the checks that ask. Signed in to raea on Pro or Studio, the app pays from your plan's monthly pool first (docs/ask.feature)
RAEA_AGENT_KEY, RAEA_API An agent key with the ask scope, and the API address it came with: asks from your plan's pool with no app running (a pipeline, CI, a server)
RAEA_ASK_ROUTE app, agent or own: ask only that way, rather than the first that answers
RAEA_ASK_DIR Where raea ask keeps its answers, the clips it cut and its cache (~/.raea/ask)
RAEA_CLEF_REPLAY A JSON file of answers by question id, given back for every call: for tests, which then never spend
RAEA_GUIDES A folder of craft references for the agent's guide tool (captions.md, motion-rules.md, ...)

Rendering

Setting What
WGPU_BACKEND The GPU driver API raea draws with: dx12, vulkan, gl (or several, comma-separated). Unset, raea picks: on Windows, DirectX 12 when the PC has an Intel or AMD GPU (also beside an NVIDIA one: their OpenGL and Vulkan drivers crashed raea there), else wgpu's own pick, which is Vulkan on an NVIDIA-only PC. Set it when the window or a render fails on a GPU, e.g. WGPU_BACKEND=dx12
RAEA_VELLO_BUMP=full Pages are drawn with Vello's own buffer sizes from the first frame, instead of raea's smaller ones (about 100 MB less GPU memory). A page past raea's sizes is found and drawn again at Vello's anyway, with a line on stderr, so this only saves that redraw
RAEA_PAGE_DRAWN_CHECK=0 Pages are drawn without checking that Vello drew them. For measuring what the check costs (about 2% in one test): a page past Vello's buffers then shows the frame before it again, or nothing
RAEA_NOT_APPLIED=0 A page's CSS is parsed without noting what stylo drops, so a render ends without its NOTE of what the page didn't apply (and its manifest's not_applied lists only loads that failed). For measuring what the note costs
RAEA_PANIC_FAULT=N A panic while frame N draws, for checking what a failing render keeps (bench/frame-panic). The render then ends with an error naming the frame and the panic, and keeps the frames drawn before it