v1.34.0
📦 weaviateView on GitHub →
✨ 6 features🐛 15 fixes🔧 7 symbols
Summary
This release introduces significant new features including Server-side Dynamic Batching and the SPFresh vector index, alongside performance improvements and dependency security updates. It also adds support for new generative AI modules and quantization methods.
✨ New Features
- Added support for RQ (Rotational Quantization) in Flat Index (1Bit + 8Bit).
- Introduced Beta Server-side Dynamic Batching.
- Added SPFresh as a new vector index, including support for virtual FS, dynamic posting size calculation, HNSW centroid index, uncompressed vectors, custom configurations, and disk queue.
- Added Contextual AI Generative and Reranker module.
- Added support for Amazon Nova Multimodal Embeddings model.
- Added support for newest Anthropic models in the generative-anthropic module.
🐛 Bug Fixes
- Ensured Unlock is called on all code paths of SPFresh.append.
- Fixed initialization of internal config for SPFresh.
- Renamed LSMStore to PostingStore in SPFresh.
- Validated sharding state integrity before use in schema-reader.
- Used unique java_outer_classname in file_replication.proto.
- Added gse_ch validatePropertyTokenization case #9340.
- Cast Bavail to uint64 to resolve Unix build error.
- Enhanced WriteTo method to ensure scratch space is cleaned up and synced after writing.
- Updated WriteTo method to return written bytes and handle scratch space cleanup errors.
- Fixed BQ -> RQ path for dynamic index upgrade.
- Ensured RoaringSet is the default strategy for dimensions bucket.
- Ensured `rf=3` single-tenant collections search locally in parallel.
- Avoided N+1 problem when calculating first strong hash.
- Reduced lock contention in `vectorCachePrefiller`.
- Passed worker ctx for retry to respect worker ctx timeout during Replication:Pull.
🔧 Affected Symbols
Flat IndexSPFreshgenerative-anthropic moduleschema-readerfile_replication.protoWriteTo methodvectorCachePrefiller