From the source
pgvector 0.7.0 introduces float16 (half-precision) HNSW indexes that consume half the memory of float32 indexes, along with sparse vector and bit vector representations.
- The update also adds L1, Hamming, and Jaccard distance functions.
- Float16 indexes can be built on existing float32 tables without changing the underlying data, or on float16 tables for additional disk space savings.
