I'm testing Seldon Core and am wondering if there is an easy way to route the output of model A as the input of model B on Seldon Core. What would be the best practice?
In a SeldonDeployment resource definition you have a graph section. The top-level element is the first step and it contains children for a next step. Here is a snippet from an example showing a transformer being used to transform a request before passing it on to a model:
graph:
children:
- children: []
implementation: XGBOOST_SERVER
name: model
modelUri: s3://model-image
type: MODEL
endpoint:
type: REST
name: feature-transformer
type: TRANSFORMER
There's an example in the seldon documentation where steps such as clean and tokenize are used to perform inference for natural language processing. That example has multiple steps as type MODEL. There's also discussion on this in the seldon slack channel.