Overview

What GQI is

GQI learns text-to-number and mixed-input-to-number mappings from labeled examples. Typical inputs include code, SQL, logs, tickets, documents, and semi-structured records.

The hosted interface is fit(X, y) followed by predict(model=fit_result.model, X=X_new). An intentional zero-shot call uses predict(model="GQI", X=X_new). Targets may be positive, zero, or negative.

Suitable tasks

• Runtime, latency, cost, or resource prediction from code and job context.
• Cardinality or outcome estimation from SQL, plans, and logs.
• Scores, durations, or costs from documents and operational records.
• Continuous outcomes from mixed text and metadata.

Purely tabular data may be better served by a tabular specialist. GQI is most useful when flattening the input into a fixed feature table would discard meaningful information.

Validate honestly

• Make the holdout match deployment. Use temporal or grouped splits when random rows would leak future or same-entity information.
• Compare against strong task-appropriate baselines.
• Report metrics such as R², Spearman, and MAE according to the target.
• Validate sample spread against held-out error before using it operationally.
• Inspect every row's token count and truncation flag.

Hosted and managed

The API exposes one model name, GQI, and private fitted state per task. Operations are asynchronous, idempotent, tenant-isolated, bounded by public alpha limits, and retained for 30 days.

Start with the Python SDK, or use the REST API reference for the exact wire contract.