pub const DEFAULT_EMBEDDING_DIM: usize = 64;Expand description
Default embedding vector dimensionality (v1.0.79, G42/S1).
Lowered from 384 to 64: with the LLM-only backend (v1.0.76+) each float
costs ~8 autoregressive output tokens, so 384 dims ≈ 3072 tokens per
vector at 50-100 tokens/s (30-60s per vector). 64 dims retain 90%+
retrieval quality for corpora under 100k memories (Matryoshka
Representation Learning, arXiv 2205.13147) while cutting generation
time ~6x. The historical 384 value matched multilingual-e5-small.