Install to your first real request

Install the SDK#

Create a fresh environment and install the released wheel with the existing Parquet reader extra. You don't need Git:

python -m venv .venv
# macOS/Linux:
. .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install "welt-client[parquet]"

Check your credits#

Open Workspace → Billing and check available credits, pending reservations and spend caps before you run the script.

Run the first request#

Download first_request.py from this reviewed documentation revision, then run it:

python first_request.py

The script connects through your browser or its printed verification link. You approve your workspace and permissions there, so you don't need to copy a key on this default path. If you already have a valid configured or saved key, the script reuses it. It never prints an API secret or writes one into a notebook or source file.

This is a real native request, not a mock or a metadata ping. The script:

• downloads the current Iris table (150 rows) and verifies it;
• uses all four features and the target column;
• splits the rows with an explicit stratified recipe (seed 9) into 120 training rows and 30 query rows;
• prepares the qualified tabicl-v2 classification configuration and returns 30 predictions within that task's declared envelope;
• prints the DataFrame and its schema, compares predictions with the held-out labels, and reports accuracy plus per-class precision, recall and F1.

These measurements describe your run. They are not a source-provided split or a publisher benchmark.

See it in your workspace#

Return to your signed-in workspace. Onboarding checks for a successful native prediction that the server attributes to your account, workspace and SDK key.

These don't count: an unrelated success in a shared workspace, a metadata GET, a queued or failed job, or a client flag. The celebration follows the server's acknowledgement, with a reduced-motion alternative.

If the operation succeeds but the console is temporarily unavailable, reconnect later. The server evidence remains durable.