Ash attribute type for a typed tensor of rank 1 to 3 (#309), backed by Nx.Tensor.

The flat row-major encoding with a declared shape covers vectors, matrices and 3-tensors; rank > 3 is not supported this release.

The Elixir-side value is an Nx.Tensor — so the value-blind structural ops (Nx.transpose/2 (lazy), Nx.reshape/2, Nx.slice/3, Nx.stack/2, Nx.concatenate/2, Nx.gather/2, …) come from Nx rather than being reimplemented here. cast_input/2 also accepts a nested list ([[1, 2], [3, 4]]) or a flat list with an explicit shape: constraint.

Storage

Neo4j has no nested-list property, so a tensor stores as a bare flat value (row-major), chosen by store:

• :property (default) — a native LIST<INTEGER|FLOAT> via Nx.to_flat_list/1; Cypher-visible.
• :packed — a base64 STRING of Nx.to_binary/1; an opaque operand.

Neither type nor shape is stored — both are declared constraints (schema, like an embedded-JSON attribute's target struct), recovered on read. So the on-disk form is just the values, and re-declaring type/shape needs no data rewrite.

Constraints

• :type — Nx element type, one of a closed set of shorthands (:u8, :s64, :f32, :f8, :c64, …). Defaults to :u8. Declared, never inferred — input is cast to it, read reconstructs from it.
• :shape — required tensor shape (e.g. [9, 9]), rank 1 to 3. Declared schema, used on write and read; never stored or inferred (no honest default exists).
• :store — :property (default) or :packed.

Usage

attribute :weights, AshNeo4j.Type.NxTensor,
  constraints: [type: :f32, shape: [3, 3]]