[Feature Request] Support for Chained/Sequential Pre-labeling Models

[chang-wd]

For example, in document analysis workflows:

Model 1 (OCR): Extract text from images → outputs text regions with bounding boxes Model 2 (Text Classification): Classify extracted text → outputs text categories (title, content, contact info, etc.) Model 3 (NER/Sentiment): Further process classified text → outputs entities or sentiment labels Currently, we have to either:

Combine all models into a single ML backend (complex and hard to maintain) Run models separately and manually combine results (inefficient) Use external preprocessing (loses the benefits of Label Studio's pre-labeling workflow)

Describe the solution you'd like Native Pipeline Support Add native support for model chaining in Label Studio with configuration like:

ml_backends:
  - name: "ocr_model"
    url: "http://localhost:9090"
    output_to: ["text_classification_model"]
  
  - name: "text_classification_model" 
    url: "http://localhost:9091"
    input_from: ["ocr_model"]
    output_to: ["sentiment_model"]
    
  - name: "sentiment_model"
    url: "http://localhost:9092" 
    input_from: ["text_classification_model"]

Describe alternatives you've considered

Additional context

[heidi-humansignal]

Hello,

Thank you for contacting Label Studio,

We truly appreciate your input and insights, I will make sure to pass along the request to our engineering team for further consideration!

Thank you again!

Comment by Oussama Assili Workflow Run