Easysearch 2.4.0: Native Vector Search in Kernel, k-NN Plugin No Longer Needed

INFINI Easysearch 2.4.0 has been officially released. The headline feature is the integration of vector search directly into the kernel, removing the requirement for an external k-NN plugin.

Native Vector Search Based on Lucene HNSW

The new version implements HNSW vector indexing natively on top of Lucene. Users can now create indexes with dense_vector fields, use query-level knn clauses, and execute top-level knn queries out of the box.

Supported vector dimensions range from 1 to 4096, covering the output of most mainstream embedding models.

Four similarity algorithms are available: cosine, dot_product, l2_norm, and max_inner_product, allowing selection per use case.

For finer tuning of index accuracy and query performance, the build parameters m and ef_construction and the query-time parameter num_candidates are exposed. Defaults are hnsw with m=16 and ef_construction=100, which work well without manual adjustment.

Hybrid Keyword and Vector Search

Version 2.4.0 introduces hybrid retrieval: a top-level keyword query and a top-level knn query can run in parallel. Their result sets are merged via a union and re-scored with weighted scoring, fusing exact keyword matching and semantic understanding in a single request. This is directly valuable for RAG pipelines, intelligent customer-service knowledge bases, and semantic search applications where both precision and semantic recall are required.

Elasticsearch 8.19.17 Compatibility

Compatibility has been a focus. This release aligns with Elasticsearch 8.19.17 for float HNSW mapping, the Bulk API, kNN query semantics, and wire-protocol behavior. It is also compatible with the official Elasticsearch Java client 8.19.17 and Python client 8.19.3. Existing applications built on ES clients can largely connect to Easysearch without major code changes.

Management and Operational Enhancements

The console adds a visual mapping editor, making field structures immediately clear.

Cluster settings now display transient, persistent, and default configuration layers simultaneously, simplifying configuration troubleshooting.

Lifecycle policies, searchable snapshots, keystore, and plugin management modules have been synchronized and optimized.

Stability of field-usage statistics and disk-usage analysis under high concurrency has been improved.

Summary

Easysearch 2.4.0 delivers three core signals:

Vector search is no longer an add-on but a kernel-native capability .

Keyword and semantic retrieval can be fused in a single query.

Full compatibility with the Elasticsearch ecosystem is maintained.

Teams building RAG, intelligent Q&A, or semantic search systems should evaluate upgrading to this version. For detailed API usage and compatibility boundaries, refer to the official native HNSW search documentation.