Change the embedding dimensions

memista is experimental (v0.1.x), and the stock crate hardcodes the embedding dimension to 2 as a demo default. Real embedding models produce 384, 768, 1536, or larger vectors, so this is the first thing you change.

Where the constant lives

The dimension is set in IndexOptions::dimensions inside src/lib.rs, in the load_or_create_index helper. Change 2 to your model’s output size:

let options = IndexOptions {
    dimensions: 768, // was 2 — set to your embedding model's size
    metric: MetricKind::IP,
    quantization: ScalarKind::F32,
    ..Default::default()
};

This is a rebuild, not a migration

The on-disk USearch format is fixed at index creation. Changing the dimension means the existing <database_id>.usearch file is incompatible — you rebuild the index by re-inserting from your source vectors.

Because memista is a crate first, the clean path is to fork it: vendor the crate, set the dimension, and pull in your fork. If you’re tuning recall vs. latency at the same time, this is also where you’d override USearch’s default connectivity and expansion_* parameters.

Related: persistence and the .usearch file, how it works.