Connect a model from your pipeline
Your pipeline registers the model once, then sends the rows it scores. Only scored rows and metadata arrive; the model itself is never uploaded. Once the model is registered it appears on the ML models tab and its checks run every 6 hours (see How checks run).
Before you begin
- An app in the LLM Gateway, and that app's Quilr API key. Every call carries this key.
- A model key: the short name the model is known by in Drift & Bias, for example
churn-model. It appears under the model name on the ML models tab. - The model's task, its feature schema and the protected attributes you want fairness checked on.
To see these steps in the console, open Assessments > Drift & Bias > ML models and select Add model.
Figure 1. The Add an ML model panel.
Step 1: Register the model
POST /v1/ml/models/{model_key}/register with the model version, task, feature schema and protected attributes.
Register again when you ship a new version. A new version gets its own reference window (see Baselines).
Step 2: Send scored rows
POST /v1/ml/models/{model_key}/rows with each batch of scored rows, up to 5,000 rows a call. Each call names a window:
referencerows: send once. These are what later rows are compared against.currentrows: send as the model scores.
If you never send reference rows, the model's first 7 days of scored rows become its reference.
The start of a rows request, as the console shows it:
POST /v1/ml/models/credit-risk/rows
Authorization: Bearer <your app's Quilr API key>
{
"modelVersion": "v3",
"window": "current",
"rows": [
{
"rowId": "r-1042",
"features": { "age": 41, "region": "north" },
"prediction": 0.73,
...
Select Copy example in the Add an ML model panel to copy a complete request body to start from.
Give every row a stable rowId. You use it later to explain a single prediction. Performance and fairness that depend on outcomes are computed only on rows that carry a label.
Step 3 (optional): Send datasets and training runs
Send dataset versions to /datasets and training runs to /training-runs. They add:
- Lineage on the model page: version, training run and dataset versions.
- Explanations: feature importance per version.
- Training runs, including failed retraining runs and their errors.