EmbeddingModel class abstract interface

Abstract interface for text-to-vector embedding models.

Allows a consuming database or application to accept an embedding model without taking a dependency on the FFI-heavy betto_inferencing package. The concrete implementation (OnnxEmbeddingModel) lives in this package and implements this interface.

Usage

Supply an EmbeddingModel to your database's open call when configuring vector indexes:

// Load the model using the concrete implementation in this package.
final model = await OnnxEmbeddingModel.load(
  cacheDir: '/path/to/cache',
);
// Pass it to your database or retrieval pipeline.
// await myDatabase.open(embeddingModel: model);

Contract

• embed is called once per document field during indexing and once per query. Implementations must be safe to call from the main isolate.
• The returned Float32List must have exactly dimensions elements.
• If text is longer than the model's context window, the implementation truncates and sets truncated = true in the returned record.
• Implementations must not throw on empty text. Behaviour for empty input is otherwise implementation-defined — for example, OnnxEmbeddingModel produces a real [CLS][SEP]-only embedding (not a zero vector), with truncated = false. A model with a mandatory EmbeddingKind prefix (e.g. multilingual-e5-small's "passage: " / "query: ") never actually tokenises truly empty content, since the prefix is prepended before empty text is checked.