Bug Fixes
- Addressed a bug where
ClientContextID
was missing from SearchRequest when library used withvectors
(#2014) - Fix to a memory leak in vector query path (#2023)
- Fix to issue #2027, error in parsing a certain datetime syntax (#2030)
- Fix that mitigates a race between persister's stale segment removal and index copy operations (#2032)
- Nested field mapping determination was broken (#2031)
Improvements
- Vector search requires the faiss dynamic library to be built from blevesearch/faiss@d9db66a
- Support for new data type -
vector_base64
which allows for interpreting vector float32s encoded as base64 following littleEndian byte ordering (#2012) - Vector dimensionality support up to 4096 from 2048 (#2015)
- New vector index optimization -
memory_efficient
alongside pre-existingrecall
andlatency
(#2040) - Better resource usage and efficiency in vector indexing (blevesearch/zapx#241, blevesearch/zapx#234)
- Delayed releasing of deserialized vector indexes to support better query throughput (#2016)