Releases: qdrant/qdrant
Releases · qdrant/qdrant
v1.11.3
v1.11.2
v1.11.1
Change log
Improvements
- #4855 - Allow to modify data before writing on disk for copy-on-write operations
- #4941 - Non-blocking payload index building
- #4947 - Include list of peers in telemetry
- #4961 - Support for MatchAny and Except conditions for UUID index
Bug fixes
- #4952 - Fix point deletions on mmap segment optimization.
- #4909 - Fix Max Upload Size for s3 snapshot storage
- #4929 - Fix configuration of bool and geo index params with explicit schema
- #4953 - Proper handle of UUID index in issue reporting
Web UI
- https://github.com/qdrant/qdrant-web-ui/releases/tag/v0.1.31 - Improved navigation and collection view
v1.11.0
Change Log
Features 🪄
- milestone#6 - GroupBy support for query API
- milestone#8 - Introduce UUID payload index
- milestone#2 - On-disk support for Keyword, Integer, Datatime, Float and UUID indexes
- milestone#5 - Random sampling support in Query API
- #4614 - Distribution-based Score Fusion
Improvements 🏇
- milestone#9 - Improve long-running update operations to not block reads
- milestone#7 - Data-locality performance improvements for multitenant and filterable collections
- #4664 - Performance improvement for prefetch queries
- #4403 - Persistent immutable id tracker allows faster load form disk
- #4658 - In-memory id tracker used avoid extra IO during segment optimization routine
- #4691 - Expose more CPU instruction set flags into CPU telemetry
- #4733 - Convert app and cluster info metrics from counters to gauges
- #4723 - Set RocksDB compression to LZ4
- #4688 - Improve result order preservation in replicated response resolution
- #4827 - Make scroll requests cancellable
- #4844, #4849, #4850 - Non-blocking retrieve with timeout and cancellation support
Bug Fixes
- #4661 - Fix NaN score for zeroed uint8 vectors
- #4663 - Fix payload update operation internal
- #4722 - Fix Charabia Chinese tokenization
- #4812, #4823 - Add missing API to metrics whitelist
- #4842 - Add and fix missing timeout in the retrieve APIs
Web UI Highlights 🍱
- qdrant/qdrant-web-ui#197 - Graph-based collection exploration tool
- qdrant/qdrant-web-ui#199 - Tool to check search quality in real time
v1.10.1
Change log
Bug fixes
- #4611 - Fix offset being applied twice in query API
- #4630 - Fix handling of score threshold for fusion requests in query API
- #4631 - Fix locally disabling last shard replica in some cases, making a shard unavailable
- #4653 - Fix not properly deleting points with delete by filter request right before hard crash
- #4589 - Make point counts when using memmap a bit more accurate
v1.10.0
Change Log
Features 🐯
-
#4225 - Universal query API - [Docs]
- Support search, recommendations, discovery, ordering by payload in a single request
- Perform hybrid search by merging results from different queries with fusion
- Build multi-stage queries with re-scoring to achieve better performance
-
#3684 - Support for multivectors - [Docs]
- Use late interaction (such as ColBERT) models natively in Qdrant
- Store dynamic number of vectors per point with shared payload
-
#4122 - Support for
float16
datatype for dense vectors - [Docs]- Reduce memory consumption for dense vectors by 2x
- Minimal accuracy loss for compared to full precision vectors
-
#4454, #4514 - Support for
float16
anduint8
datatype for sparse vectors index - [Docs]- Reduce memory consumption for sparse vectors by factor of 2 and 4 correspondingly
- Minimal accuracy loss for compared to full precision vectors
-
#4126 - Support for Inverse Document Frequency (IDF) in sparse vectros - [Docs]
- Provides an option to automatically account for IDF in sparse vectors search
- Unlocks the ability to stream updates for BM25 and BM42 embeddings
-
#3324 - Support for S3-compatible storage for snapshots - [Docs]
-
#3471 - Issue reporting
- Get notified about potential performance issues and misconfigurations
-
#4299 - Ability to configure default collection parameters
- Pre-configure quantization, vector storage and replication factor for new collections
-
#4317 - Ability to overwrite global optimizer configuration for collections
- Allows to separate roles for indexing and searching within the single qdrant cluster
Improvements 🦈
-
#4253, #4350 - Delta Encoding and bitpacking compression for sparse vectors
- Reduce memory consumption for sparse vectors by up to 75%
-
#4527 - Skip serialization of empty fields in the search response
- Less confusion in the search response
- Reduce network traffic
-
#4267 - Add test for OOD during indexing, prevents crash on out-of-disk
Bug Fixes 🦟
- #4553 - Fix exclusion of ids in discovery api
- #4583 - Fix potential storage corruption during hard shutdown
Highlights 💡
- Significant refactoring of the Rust Client making it more extendable, easier to use and more idiomatic
- New API reference web-page - https://api.qdrant.tech/
- New fast sparse embedding model - BM42, aimed to resolve limitations of SPLADE and fix full-text search issue for RAG
v1.9.7
v1.9.6
v1.9.5
Change log
Features
- #4254 - Add pyroscope integration for continuous profiling on demand
Improvements
- #4309 - Allow to configure default number of shards per node
- #4317 - Allow to overwrite optimizer settings via config
- #4312, #4369 - Improve vector size estimations, making index thresholds more reliable
- #4428 - Improve default maximum segment size, base it on number of CPUs used for indexing
- #4370 - Use consistent RocksDB settings for both put and remove
- #4376 - Improve ordering of insertions and deletions in RocksDB
- #4371 - Log error if segment flushing failed on drop
- #4352 - Promote REST request processing problems from warning to error
- #4368 - Improve error messages in cases of missing vectors
- #4391 - Improve shard state log message, not strictly related to snapshot recovery
- #4414 - Improve Dockerfile, don't invalidate caches each commit and allow debug settings
Bug fixes
- #4402 - Fix deadlock caused by concurrent snapshot and optimization
- #4411 - Fix potentially losing vectors on crash by enabling RocksDB WAL
- #4416, #4440 - Respect
max_segment_size
on data ingestion with optimizers disabled, create segments as needed - #4442 - Fix potentially having bad HNSW links on multithreaded systems