Python SDK

gqi-labs is the dependency-free distribution for the hosted API. It installs the gqi Python module and contains only HTTP transport and public wire types; model weights, training, serving, and cloud implementation are not distributed.

pip install gqi-labs

Interface

from gqi import GQI

client = GQI(api_key="gqi_sk_...")
fitted = client.fit(
    task_name="ticket-priority",
    X=[
        "checkout fails", "minor copy edit", "refund is stuck", "email delayed",
        "production is down", "feature request", "wrong invoice", "broken link",
    ],
    y=[10, 1, 7, 4, 10, 2, 5, 2],
)
result = client.predict(
    model=fitted.model,
    X=["payments fail for every customer"],
)
print(result.rows[0].prediction, result.rows[0].iqr)

Use client.predict(model="GQI", X=X_new) for an intentional zero-shot call. Prediction does not silently choose the base when model is omitted. LoRA is the default fit method; config={"method": "full"} returns an immutable dedicated-model handle.

fit() and predict() block until their jobs finish. Long-running callers can use submit_fit(), submit_predict(), job(), wait(), and cancel().

• Every finite target is retained.
• A fitted ModelHandle, or explicit "GQI" for zero-shot, is required by prediction.
• FitConfig and PredictConfig validate configuration before submission.
• Model and engine implementation details remain server-side.

Set GQI_API_KEY to omit api_key from the constructor. See the REST API reference for the underlying wire contract and hosted limits.