CHANGELOG

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. See the CONTRIBUTING guide for instructions on how to add changelog entries.

Unreleased 3.x

Add model_selection (language_option/model_type) parameter to semantic field to resolve the model id from cluster settings (#1918)

[Neural Query] Expose embedded filters to QueryBuilderVisitor traversal (#1992)
[Hybrid Query] Add opt-in index setting index.neural_search.hybrid_collapse_distinct_groups_enabled to make collapse return top-size distinct groups instead of deduplicated top-size documents (#1947)
[Hybrid Query] Fix NoSuchElementException in hybrid query with sort/search_after when a shard returns no results (#1939)
[SemanticHighlighter] Fix SemanticHighlighterExtBuilder.toXContent (#1906) (query-insights #651)
[Sparse ANN] Fold sparse vector tokens into the signed-short range (modulus 32768) so folded tokens are never sign-extended to a negative value when stored in short[] (#1926)
[Hybrid Query] Read the current document and its sub-query matches from the positioned disjunction iterator, fixing an ArrayIndexOutOfBoundsException and silently misattributed scores when a sub-query has a two-phase iterator (#1946)
[RRF] Reject a combination technique other than rrf when creating a score-ranker-processor, instead of accepting the pipeline and throwing NullPointerException on every query (#1949)
[Sparse ANN] Skip cache cleanup for a closed index's shards, which have no MapperService, so reopening a sparse index no longer leaks the shard lock and leaves the shard unassigned (#1982)

[Sparse ANN] Add scripts/build.sh so the distribution build ships the native sparse engine: every SIMD variant the target architecture may need is built into the plugin zip, and the variant to load is picked from the host's CPU flags at runtime (#1978)
[Sparse ANN] Add the JNI layer for the native sparse engine, bridging to the neural-sparse-cpp library, with a googletest suite run on Linux and Windows plus an ASan/LSan job (#1972)
[Neural Sparse] Pin the two-phase processor IT index to a single shard so its pruned-score assertions do not depend on the cluster's default shard count (#1959)
[Semantic Field] Add an end-to-end remote dense model IT for the semantic field mapping transformer using the TorchServe mock model (#1966)
Guard eclipse() to spotless tasks and keep P2 mirror on pinned version (#1988)

[Hybrid Query] Add cross-plugin integration coverage for document-level security (#1957)

[SemanticHighlighter] Traverse the query tree with a QueryBuilderVisitor instead of a "manual" walk (#1915)
[RRF] Compute rank scores with exact integer arithmetic instead of allocating a BigDecimal per document (#1942)