Neural Search Comes to Apache Solr: Approximate Nearest Neighbor, BERT & more
summary
Apache Solr adds first-class neural search in 9.0: store vectors in a dense vector field, index with Lucene’s HNSW proximity graphs and run Approximate KNN queries via a knn query parser. Supported distances: cosine, euclidean, dot; default practical recommendation is a 768-dim BERT embedding (Solr caps vectors at 1024). Index build time and memory rise (graph construction), but a 1–2 GB test index shows faster query latency than simple lexical search. Current limits: post-filtering and re-ranking only apply to the top-k, no docValues/multivalued vectors; pre-filtering lands in 9.1.
AI-written summary · about
chapters (30)
- 0:00Introduction
- 0:30About me
- 2:08About my company
- 3:37Overview
- 4:58Vocabulary Mismatch
- 6:24False Negative Results
- 7:39Semantic Similarity
- 8:20Vector Representation
- 10:23Indexing Time
- 11:38Distance between vectors
- 12:22Queries for vectors
- 12:50Approximate Nearest Neighbor
- 14:37Hierarchical Navigable Small World Graph
- 18:38Solr similarity functions
- 20:14Query time
- 21:49Solr schema
- 23:21Solr functions
- 24:19Advanced parameters
- 25:33Indexing vectors
- 26:43Kn query parser
- 28:50Reranking
- 29:47Hybrid dense sparse retrieval
- 31:00Benchmarks
- 31:58Generating vectors from text
- 34:26BERT
- 36:18Python
- 37:35Performance
- 38:09Prefiltering
- 39:15Vector similarity simplification
- 39:46Update request processor
related talks
28mHybrid Search with Apache Solr
41mApache Solr 10: What's Coming up for Vector Search
01h 09mVector Search: Ask Me Anything!
37mLexical, Vector & Hybrid Search with Elasticsearch
42mResearch in the Age of AI
38mSearch: A New Era
47mLatest innovations from Elasticsearch: Drive speed, scale & relevance (PRT266)
46mNatural Language & Search, Improve your Search w/ Vector capabilities