The CLI#
Talks to Embedding Studio: everything the console does, JSON out, non-zero exit on refusal, which is the contract a script or a coding agent needs.
# global, any OS, Node 18+
npm install -g embedding-studio-cli
# or one Python 3.8+ file, no pip
curl -fsSL https://studio.embedding.in/cli.py -o studio && chmod +x studio
# Windows, PowerShell
iwr -useb https://studio.embedding.in/cli.py -OutFile studio.pyAuth is STUDIO_API_KEY in the environment, issued in the console under API keys. STUDIO_API_URL points it at a self-hosted deployment.
projects | project <slug> | create | delete <slug> --yes
ask --template <t> | --questions file.json # decide, no project
upload <slug> file.csv [--watch] sample <slug> file [--watch]
job <id> [--watch] | stop <id> | restart <id>
fits <slug> | fit <id> | band <fit_id> --precision X | --review-rate Y
serving <slug> | calibration <slug> [-o file] | decide <slug> --text "..."
decisions | audit | drift --sample-id <id> | keys | billing | healthJSON on stdout, non-zero exit with the API's message on stderr: the contract a script or a coding agent can build on.
The engine#
pip install vegamlPython 3.10 or newer, on Linux, macOS or Windows. Dependencies: torch, transformers, safetensors, huggingface_hub, numpy, nothing else. The serving and calibration tools are extras: pip install "vegaml[serve]" adds vegaml-serve and vegaml-calibrate.
Watch out
Check which Python pip points at.
macOS ships a system python3 at 3.9, below the floor, and the error is a resolver complaint rather than a version message. python3 -V first, then python3.12 -m venv .venv && . .venv/bin/activate.
Readers, their disk and memory costs, and device selection: Readers.