Skip to main content

Design Decisions

info

This page is coming soon. Below is an outline of what it will cover.

Planned Content​

  • Why a facade pattern — simplicity, discoverability, and a flat API surface
  • Why Protocol classes over ABC — structural subtyping and no inheritance requirement
  • Why DataFrames for the IR — familiar to data scientists, interoperable with pandas/polars
  • Why optional extras — keeping the core lightweight, installing only what you need
  • Why not a plugin system — tradeoffs of registry vs plugin discovery
  • Why calibrate against PowerGraph — NeurIPS 2024 benchmark credibility