Change Log

Features :tiger:

• https://github.com/qdrant/qdrant/issues/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

• https://github.com/qdrant/qdrant/issues/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

• https://github.com/qdrant/qdrant/pull/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

• https://github.com/qdrant/qdrant/pull/4454, https://github.com/qdrant/qdrant/pull/4514 - Support for float16 and uint8 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

• https://github.com/qdrant/qdrant/pull/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

• https://github.com/qdrant/qdrant/issues/3324 - Support for S3-compatible storage for snapshots - [Docs]

• https://github.com/qdrant/qdrant/issues/3471 - Issue reporting

  • Get notified about potential performance issues and misconfigurations

• https://github.com/qdrant/qdrant/pull/4299 - Ability to configure default collection parameters

  • Pre-configure quantization, vector storage and replication factor for new collections

• https://github.com/qdrant/qdrant/pull/4317 - Ability to overwrite global optimizer configuration for collections

  • Allows to separate roles for indexing and searching within the single qdrant cluster

Improvements :shark:

• https://github.com/qdrant/qdrant/pull/4253, https://github.com/qdrant/qdrant/pull/4350 - Delta Encoding and bitpacking compression for sparse vectors

  • Reduce memory consumption for sparse vectors by up to 75%

• https://github.com/qdrant/qdrant/pull/4527 - Skip serialization of empty fields in the search response

  • Less confusion in the search response
  • Reduce network traffic

• https://github.com/qdrant/qdrant/pull/4267 - Add test for OOD during indexing, prevents crash on out-of-disk

Bug Fixes :mosquito:

• https://github.com/qdrant/qdrant/pull/4553 - Fix exclusion of ids in discovery api
• https://github.com/qdrant/qdrant/pull/4583 - Fix potential storage corruption during hard shutdown

Highlights :bulb:

• 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