Extend ML documentation

docs/integration/MachineLearning.md

@@ -4,9 +4,46 @@ sidebar_position: 5
 
 # Machine Learning (Enterprise only)
 
-Nussknacker can infer ML models using the Machine Learning enrichers. The ML enrichers are Enterprise components of Nussknacker and require a separate license. Please contact <[email protected]> for license terms and more details.
+Nussknacker can infer ML models using the Machine Learning enrichers. We support both in-process inference and inference
+using a dedicated ML runtime.
 
-We support the inference of the following ML technologies:
-- native Python models discovered using the [MLflow](https://mlflow.org/) model registry and executed with our ML runtime
-- models exported in the [PMML](https://en.wikipedia.org/wiki/Predictive_Model_Markup_Language) format
-- models exported in the [H2O](https://h2o.ai/) format
+Nussknacker ML enrichers can infer **in-process** models exported in the following formats:
+- [ONNX](https://onnx.ai/)
+- [PMML](https://en.wikipedia.org/wiki/Predictive_Model_Markup_Language)
+- [H2O](https://h2o.ai/)
+
+Executing ML models in the same process as the Nussknacker scenario is a matter of configuring an ML enricher. No 
+additional components are required to infer models. Also, invoking exported models using their native runtimes 
+should result in both minimal latency and low resource requirements. However, this approach has some limitations. 
+Namely, not all models and not all popular ML libraries are covered by these formats. The export process itself
+can feel cumbersome.
+
+Nussknacker can discover exported models from file (e.g. local directory, NFS) or HTTP based registries.
+We also plan to support MLflow as a registry for exported models in these formats.
+
+The diagram below shows the interactions for ML enricher inferring an exported model:
+
+![alt_text](img/mlExportedModels.png "Inferring ML exported models")
+
+Because of some restrictions with exporting models, we also support inference of **any Python ML model** using our
+*Nussknacker ML runtime* component. ML models are discovered and fetched from the [MLflow](https://mlflow.org/)
+model registry which also serves as an experiment tracking tool for data scientists. This way, data scientists do not
+have to worry about exporting trained models to other formats, they just log the models in the MLflow model registry.
+They also have the flexibility to use various Python ML libraries.
+
+Nussknacker ML runtime offers higher throughput at the cost of higher inference latency. Since the ML runtime
+is a separate component of Nussknacker installation, a GPU can be also harnessed to infer ML models.
+
+The diagram below shows the interactions for MLflow enricher and ML runtime: 
+
+![alt_text](img/mlMlflowAndNussknackerMlRuntime.png "Inferring MLflow models with Nussknacker ML runtime")
+
+A Kubernetes cluster is recommended for installing the Nussknacker ML runtime. However, the ML runtime can also be
+installed as a standalone application.
+
+You can read more about this approach in our [blog post](https://nussknacker.io/blog/ml-models-inference-in-fraud-detection/)
+where we present a complete path from analysing a dataset to training a model and finally inferring the model
+in a Nussknacker scenario.
+
+The ML enrichers are Enterprise components of Nussknacker and require a separate license. Please contact
+<[email protected]> for license terms and more details.

docs/scenarios_authoring/Enrichers.md

@@ -152,15 +152,19 @@ Similarly, information about field names and types returned by the OpenAPI servi
 ## ML enricher
 **(Enterprise only)**
 
-Nussknacker can infer ML models using the Machine Learning enrichers. The ML enrichers are Enterprise components of Nussknacker and require a separate license. Please contact <[email protected]> for license terms and more details.
+Similarly to SQL and OpenAPI enrichers, as ML model input and output vectors are known to Designer, when you double-click 
+the ML Enricher node in the scenario you will see entry fields required by the selected model and version;
+data type hints and syntax error checking functionality will be active.
 
-We support the inference of the following ML technologies:
-- native Python models discovered using the [MLflow](https://mlflow.org/) model registry and executed with our ML runtime
-- models exported in the [PMML](https://en.wikipedia.org/wiki/Predictive_Model_Markup_Language) format
-- models exported in the [H2O](https://h2o.ai/) format
+From the scenario author perspective, the ML Enricher is indistinguishable from OpenAPI enricher - it just takes 
+some input parameters and returns a value, e.g. a risk.
 
-Similarly to SQL and OpenAPI enrichers, as ML model input and output are known to Designer, when you double-click the ML Enricher node in the scenario you will see entry fields required by the model; data type hints and syntax error checking functionality will be active.   
+![alt_text](img/mlEnricherForm.png "ML Enricher")
 
-From the scenario author perspective, the ML Enricher is indistinguishable from OpenAPI enricher - it just takes some input parameters and returns a value. 
+The following ML models are supported:
+- [ONNX](https://onnx.ai/)
+- [PMML](https://en.wikipedia.org/wiki/Predictive_Model_Markup_Language)
+- [H2O](https://h2o.ai/)
+- models fetched from [MLflow](https://mlflow.org/) models registry
 
-![alt_text](img/mlEnricherForm.png "ML Enricher")
+Please check [ML integration](../integration/MachineLearning.md) page for more details on how models are inferred by Nussknacker.