opensearch-knn
Installation
SKILL.md
OpenSearch k-NN
What Changed at 2.13+
OpenSearch 2.13 through 2.17 made the k-NN plugin competitive with purpose-built stores:
- Lucene engine supports HNSW with filter-while-search (efficient filtering).
- Faiss engine added IVF + PQ + binary + int8 quantization.
- Neural search plugin: server-side embedding via external models (SageMaker, Bedrock).
- Radial search: return all neighbors within a max distance.
- Hybrid search plugin: BM25 + kNN with configurable score combiners.
Engines: nmslib vs faiss vs lucene
| Engine | Algorithms | Pros | Cons |
|---|---|---|---|
| lucene | HNSW | Best filter-while-search, pure-Java | No quantization, no IVF |
| faiss | HNSW, IVF, IVF_PQ, binary | Quantization, scale, GPU path | Filters are post-filter by default |
| nmslib | HNSW | Legacy default | Being phased out; no new features |