Database

DashAI uses SQLite as its database (stored at ~/.DashAI/db.sqlite) with SQLAlchemy as ORM and Alembic for schema migrations.

Key Tables

Table	Purpose
Dataset	Uploaded dataset — name, Arrow file path, loading status, and timestamps.
ModelSession	Experiment configuration — dataset, task name, input/output columns, train/validation/test split ratios, and selected metrics per split.
Run	Individual training execution within a ModelSession — model name, parameters, optimizer config, goal metric, run artifacts, execution status and timing, and paths to optimization plots (history, slice, contour, importance).
Metric	Single metric measurement — name, value, split (TRAIN/VALIDATION/TEST), level (LAST/STEP/BATCH/TRIAL), and step index. Linked to a Run.
Prediction	Prediction job — links a trained Run to an input Dataset, tracks execution status and timing, and stores the path to output results.
GenerativeSession	Generative model session — task type, model name, current parameters, and a human-readable name and description. Owns a history of parameter snapshots and all associated GenerativeProcess records.
GenerativeProcess	Single invocation of a GenerativeSession — tracks execution status and timing. Linked to ProcessData records that hold the input and output payloads.
ProcessData	Input or output payload for a GenerativeProcess — serialized data value, data type (text, image, etc.), and an is_input flag to distinguish inputs from outputs.
GenerativeSessionParameterHistory	Immutable snapshot of a GenerativeSession's parameters captured at each change, providing a full audit trail of parameter evolution over time.
Notebook	Working dataset session — a mutable copy of a source Dataset on which Explorers and Converters can be applied. Changes can be reverted; the result can be saved as a new Dataset for model training.
Explorer	Visualization record within a Notebook — explorer type, selected columns, parameters, path to saved results, and execution status.
Converter	Single converter step applied to a Notebook's mutable dataset — converter type, parameters, execution status, and timing. Multiple records form an ordered transformation pipeline on the Notebook.
Plugin	Installed plugin — name, author, installed and latest versions, status, summary, and full description. Owns Tag records for classification.
Tag	Classification tag for a Plugin (e.g., Model, Task, Metric), used for filtering and discovery.
GlobalExplainer	Global model explanation — explainer type, linked Run, parameters, paths to explanation data and plot, and execution status. Covers the model as a whole.
LocalExplainer	Local (per-instance) explanation — explainer type, linked Run and Dataset, parameters, fit parameters, scope, result paths, and execution status.

Important Enums

• RunStatus: NOT_STARTED → DELIVERED → STARTED → FINISHED | ERROR
• SplitEnum: TRAIN, VALIDATION, TEST
• LevelEnum: LAST (final value), STEP, BATCH, TRIAL (for optimization)

Data Storage

• Datasets are stored in Apache Arrow IPC format (columnar, efficient for ML workloads).
• Trained models are saved as pickle/joblib files under ~/.DashAI/runs/{run_id}/.
• Plots generated during hyperparameter optimization are stored as serialized Plotly objects.
• Metric time-series (per step, batch, or trial) are stored in the Metric table for tracking training progress.