Fun fact: We've implemented binary embedding search [1] without the need for a specialized vector database. Instead, dimensional tokens like 'embedding_0_0', 'embedding_1_0' are created and being built into the tantivy index [2].<p>We're satisfied with the quality and performance this approach yields, while still keep Tabby embed everything into a single binary.<p>[1] My binary vector search is better than your FP32 vectors: <a href="https://blog.pgvecto.rs/my-binary-vector-search-is-better-than-your-fp32-vectors" rel="nofollow">https://blog.pgvecto.rs/my-binary-vector-search-is-better-th...</a><p>[2] Tantivy: <a href="https://github.com/quickwit-oss/tantivy">https://github.com/quickwit-oss/tantivy</a>