On 3rd March 2026, the Solr PMC announced the release of Apache Solr 10.0.0!
Solr 10.0.0 is now available for immediate download here.
In this blog post, we summarise the main new features introduced in this major release, focusing on Vector Search and Learning to Rank (LTR).
These features were contributed by the Sease Ltd team (specifically by Alessandro Benedetti, Anna Ruggero, and me, Ilaria Petreti), as part of our ongoing effort to enhance Solr’s search and ranking capabilities, and by Kevin Liang, a senior software engineer at Bloomberg.
Below, we will provide a brief overview of each new feature, followed by a link to the full blog post where each topic is covered in more detail.
Vector Search
1) Add Support for Scalar Quantized Dense Vectors
Jira Issue: SOLR-17780
Solr GitHub PR: #3385
A new schema field type, named ScalarQuantizedDenseVectorField, has been introduced. It enables scalar quantization for vector data, which reduces the precision of floating-point values to a smaller number of bits — in this case, either 4 or 7 (default) — therefore reducing memory consumption and speeding up search for vector indexes.
2) Add Support for Binary Quantized Dense Vector
Jira Issue: SOLR-17812
Solr GitHub PR: #3468
A new schema field type, named BinaryQuantizedDenseVectorField, has been introduced. It enables binary quantization for vector data, which reduces the precision of floating-point values to a single bit.
Similar benefits to scalar quantisation (with a more extreme trade-off)
3) Add Support for PatienceKnnVectorQuery (Early Termination Strategy)
Jira Issue: SOLR-17814
Solr GitHub PR: #3644
PatienceKnnVectorQuery is a version of knn vector query that exits early the graph when HNSW queue saturates over a saturationThreshold for more than patience times. With this support, we have introduced an early termination strategy for knn queries that makes approximate k-NN search faster and more efficient.
4) Add Support for SeededKnnVectorQuery
Jira Issue: SOLR-17813
Solr GitHub PR: #3705
SeededKnnVectorQuery is a version of knn vector query that brings a powerful enhancement to vector search capabilities. With this support, users can now perform k-nearest-neighbors (kNN) searches that start from a specific “seed” query, to narrow down the candidate set before performing the vector similarity computation. The “seed” query is primarily intended to be a lexical query, guiding the vector search in a hybrid-like way through traditional query logic, i.e. documents matching the seed query will be used as entry points in the HNSW graph exploration.
5) ACORN-Based Filtering in Vector Search
Jira Issue: SOLR-17815
Solr GitHub PR: #3680
ACORN (introduced in this paper) is an algorithm that presents an optimized solution for filtered vector search.
This approach addresses the performance limitations typically associated with pre-filtering and post-filtering strategies by modifying both the construction and search phases of the HNSW graph.
With this contribution, Solr now supports the ability to pass the Lucene filteredSearchThreshold parameter directly as a KNN query parameter. This threshold determines when the ACORN optimization is applied: if the percentage of documents satisfying the filter falls below the specified threshold, the ACORN approach is automatically used.
6) Rename Vector Search Parameters and Language Model Modules
Jira Issue: SOLR-17927
Solr GitHub PR: #3789
The neural package has been renamed to vector and the llm module has been renamed to language-models. This change is purely a code-level refactoring and does not affect end users.
The HNSW parameters hnswMaxConnections and hnswBeamWidth have been renamed to hnswM and hnswEfConstruction, respectively. These changes do affect users, who must update the corresponding entries in their schema.xml file when upgrading Solr.
However, fallback support for the legacy names (hnswMaxConnections, hnswBeamWidth) has been included to reduce friction during upgrades and maintain backward compatibility.
Learning to Rank
Add Feature Vector Caching for Ranking
Jira Issue: SOLR-16667
Solr GitHub PR: #3433
Solr already includes a cache mechanism for Learning to Rank feature vectors; however, it is limited to the feature logging phase and not used during reranking. Our contribution extends this capability by introducing a new Learning to Rank cache that Solr leverages in both phases, significantly improving search performance.
In our blog post, you will find a detailed explanation of the new cache, along with a benchmark comparing it to the previous implementation, which confirms the significant performance improvements achieved with the new cache.
Power the Future of Solr
Power the future of Solr
Contribute or sponsor! Every pull request, bug report, documentation improvement, and donation helps the community deliver faster fixes, stronger features, and more reliable releases for everyone.
