Running Predictions with HuggingFace

Once you have a trained model — or want to use any public HuggingFace model — you can run inference directly from XGENIA.

What you will learn in this guide

• How to run predictions through the AI chat
• How to use the Auto ML Predictor pro node
• Serverless vs. dedicated endpoint inference

Inference Modes

Mode	Description	Best For
Serverless	Uses HuggingFace's free inference API	Testing, low-volume predictions
Dedicated Endpoint	Uses a paid, always-on endpoint	Production workloads, guaranteed uptime

Predictions via AI Chat

Ask the AI to run a prediction with any HuggingFace model:

Classify this text using bert-base-uncased: "I love this product, it works great!"

Or use a custom model you trained:

Predict churn for this customer using my-username/churn-model:
{"tenure": 12, "monthly_charges": 65.5, "contract": "month-to-month"}

The AI calls hf_predict under the hood, which sends the request to HuggingFace's Inference API.

Response Format

{
  "success": true,
  "prediction": [{"label": "POSITIVE", "score": 0.998}],
  "model": "bert-base-uncased",
  "mode": "serverless"
}

Using the Auto ML Predictor Pro Node

For visual workflows, use the Auto ML Predictor node.

Inputs

Port	Type	Description
inputData	Any	Data to predict on (text, object, or array)
modelRepo	String	HuggingFace model repository (e.g. username/my-model)
inferenceMode	Enum	serverless or endpoint
endpointUrl	String	Dedicated endpoint URL (only for endpoint mode)
parameters	Object	Model parameters (e.g. {temperature: 0.7})
hfToken	String	HuggingFace API token
mlServerUrl	String	ML Coordinator URL
Predict	Signal	Trigger prediction

Outputs

Port	Type	Description
prediction	Any	The model's prediction result
confidence	Number	Confidence score (0–1)
rawResponse	Object	Full response from HuggingFace
error	String	Error message if prediction fails

Tips

• Cold start: Serverless models may take 10–30 seconds on first call while the model loads
• Rate limits: Free tier has rate limits — use dedicated endpoints for production
• Model compatibility: Ensure your input format matches what the model expects
• Parameters: Use max_length, temperature, top_p for text generation models

End-to-End Example

A typical ML pipeline on the canvas:

1. Query Records → fetches customer data from your database
2. Auto ML Analyzer → analyzes the data, suggests target and features
3. Auto ML Trainer → trains a model on HuggingFace
4. Auto ML Predictor → runs predictions with the trained model
5. UI components → display predictions to users

Wire the Trainer's hfModelRepo output directly to the Predictor's modelRepo input for seamless end-to-end workflows.