Agnostic data ingestion

“Be liberal in what you accept.” Every surface takes whatever data you already have and normalises it once to Polars internally — no reshaping, no manual schema.

list[dict]

from dash_tensor_grid import GridDataEngine

engine = GridDataEngine([
    {"region": "EMEA", "revenue": 100},
    {"region": "AMER", "revenue": 200},
])

Polars

import polars as pl
from dash_tensor_grid import GridDataEngine

engine = GridDataEngine(pl.read_parquet("sales.parquet"))   # DataFrame or LazyFrame

pandas

import pandas as pd
from dash_tensor_grid import GridDataEngine

engine = GridDataEngine(pd.read_csv("sales.csv"))            # needs the [pandas] extra

JSON string

from dash_tensor_grid import GridDataEngine

engine = GridDataEngine('[{"region": "EMEA", "revenue": 100}]')

In the browser

The framework adapters take a plain array of row objects (Record<string, unknown>[]) — parse your JSON/CSV however you like and pass the array straight in; the engine port does the rest.

mountTensorGrid(el, { data: rows, rows: ['region'], measures: { revenue: 'sum' } });

Numeric correctness

The engine aggregates raw numerics (never stringify-then-sum), preserves precision, and serializes NaN / Inf / None deliberately as null so the frontend renders — instead of crashing JSON.parse. Dates become epoch-ms / ISO-8601; exact money is scaled/Decimal-aware on the ingestion path.