Skip to content

Data Science Overview

Flow-Like combines data access, SQL, machine-learning nodes, model inference, and A2UI visualization in one workflow environment. Use Flows to make the analysis steps and operational handoff inspectable.

The Flow-Like data science workflow: load data, explore and transform it, analyze it with machine learning or GenAI, visualize findings, and operationalize the result

ApplicationFlow-Like role
Data explorationRegister sources, inspect schemas, and query with SQL
Repeatable analysisTurn cleaning, feature preparation, and metrics into a Flow
Machine learningSplit datasets, scale and vectorize features, train supported models, tune them with grid search or AutoML, evaluate, inspect, save, and predict
Model inferenceRun ONNX vision, OCR, audio, NLP, and related models
Interactive dashboardsPush query results into A2UI tables and charts
AI-assisted analysisCombine governed data tools with configured models or agents

Flow-Like can work with local files, databases, app storage, APIs, and data-lake tables. Start with Data loading and storage when choosing the reader and storage boundary.

DataFusion entry points include:

Create DataFusion Session provides the query context. Register the required sources, inspect them with List Tables and Describe Table, then use SQL Query.

See DataFusion and SQL for joins, aggregations, and federated analysis.

The generated node catalog currently includes:

TaskExamples
Dataset preparationSplit Dataset, Stratified Split, K-Fold Split, Shuffle Dataset, Sample Dataset
PreprocessingFit Feature Scaler, Fit TF-IDF Vectorizer, Apply Transform
ClassificationDecision Tree, Naive Bayes, Multinomial Naive Bayes, Logistic Regression, SVM, K-Nearest Neighbours, Random Forest, AdaBoost, One-Class SVM
RegressionLinear, Ridge/Lasso/ElasticNet, GLM / Tweedie, K-Nearest Neighbours, SVM
Ordinal regressionProportional Odds, Ridge, Frank & Hall, Continuation Ratio, Adjacent Category, Neural CORAL/CORN
ClusteringKMeans, DBSCAN, Gaussian Mixture
ReductionPCA, t-SNE
EvaluationAccuracy, Confusion Matrix, Regression Metrics, ROC-AUC & Log Loss, Silhouette Score, Ordinal Metrics
Tuning and AutoMLGrid Search, Auto Classifier, Ordinal Grid Search, Auto Ordinal
Model introspectionModel Info, Feature Importance, Get Coefficients, Get Centroids
Persistence and inferenceSave Model, Load Model, Save Model (Binary), Load Model (Binary), Predict, Apply Transform

Predict serves the trainers that emit a model. DBSCAN, PCA, and t-SNE return cluster counts, explained variance, or a written-back embedding column instead of a model, and fitted transformers are replayed with Apply Transform rather than Predict.

See Machine learning for the end-to-end workflow, Advanced configuration for the knobs on each trainer, and Auto training for the tuning nodes.

The ONNX node catalog includes tasks such as image classification, object detection, segmentation, OCR, face processing, pose estimation, audio preparation, voice activity detection, and named-entity recognition. Model compatibility still depends on the expected input and output tensors; inspect model metadata and validate preprocessing before relying on predictions.

Use A2UI NivoChart or PlotlyChart components for interactive analysis and Table for detailed records. Workflow nodes can push query results into charts and tables.

See Data visualization for chart selection and Building internal tools for page, data, and action wiring.

A compact sales workflow has four explicit stages:

  1. mount the CSV or register the source table;
  2. validate the schema and relevant time range;
  3. aggregate revenue by region with SQL;
  4. push the result to a bar chart and a supporting table.

The table lets a reviewer verify the values behind the chart, while the workflow preserves the query used to calculate them.

  • Source versions or query windows are recorded
  • Schema and data-quality checks run before analysis
  • Random seeds are fixed where supported
  • Train, validation, and test data remain separate
  • Preprocessing used for training is reused for inference — a fitted Feature Scaler replays its exact offsets and scales, while TF-IDF recomputes its document frequencies inside each Apply Transform run
  • Model and metric configuration is versioned
  • Results include units, filters, and known limitations
  • Sensitive features and outputs follow the app’s access policy