Benchmarks
Measured on the industry's own benchmarks.
Up to 80x faster than Elasticsearch, 31x faster p99 than Amazon S3 Vectors and 2.7x faster p99 than turbopuffer. Each result comes from the benchmark that vendor publishes against: Elastic's Rally tracks, VectorDBBench and turbopuffer's open harness.
Headline results
Where we stand
Each number is WarpSearch against the competitor's own published figure for the same workload.
faster geo distance queries on 60.8M points
28 ms vs 2,254 ms for distance ranges on Rally's geopoint track. Bounding box, distance and polygon filters answer in 1 ms, where Elasticsearch's nightly run takes 11 to 43 ms.
Geo results → vs Elasticsearch 6.8xfaster nested aggregations on Rally
145 ms vs 984 ms p50 on the official nested track's date histogram. Nested queries are 2.9x faster and bulk indexing 2.7x faster.
Rally results → vs Amazon S3 Vectors 31xfaster p99 at the same recall
11 ms p99 at recall 0.88 on VectorDBBench Cohere 1M, where S3 Vectors publishes 337 ms at recall 0.87. pgvector publishes 731 ms at recall 0.89; WarpSearch takes 12 ms at 0.90.
Vector results → vs turbopuffer 2.7xfaster p99 on turbopuffer's own 10M workload
turbopuffer publishes 14 ms p50 and 27 ms p99 for 10M 1024-d vectors at 8 queries a second. WarpSearch: 9 ms and 10 ms.
turbopuffer results →Benchmark families
Every kind of search, measured the way the industry measures it
We use the suites other engines already publish against, so every result can be checked against a number someone else produced.
Elasticsearch compatibility
Elastic's own Rally tracks sent to WarpSearch's Elasticsearch API: full-text, sorting, aggregations, nested documents, geo and scroll.
Compared with: Elasticsearch nightly benchmarks
See Rally results →Vector search
Recall against latency and load time for approximate nearest-neighbour search, on the dataset vector databases publish against.
Compared with: S3 Vectors, pgvector, turbopuffer, Weaviate, Elastic Cloud, OpenSearch, Qdrant
See vector results →turbopuffer workloads
turbopuffer's open harness and website configuration, run against WarpSearch's turbopuffer-compatible API.
Compared with: turbopuffer's homepage, docs and blog figures
See turbopuffer results →Percolation (reverse search)
Store millions of queries and find the ones that match each new document. On Rally's percolator track with 2M stored queries, WarpSearch is 2.7x to 5.9x faster than Elasticsearch.
Hybrid search and relevance
Ranking quality for keyword, vector and hybrid search on judged query sets, next to Elasticsearch, OpenSearch and Apache Pinot.
Logs and observability
Log ingestion, time ranges and aggregations on the log tracks Elasticsearch and OpenSearch publish against.
How we measure
Four rules every result follows
Their benchmark, their numbers.We pick workloads a competitor already publishes, run them on WarpSearch, and quote the competitor's own figure with its source and date. We do not install or tune other engines.
Official tracks and harnesses.Rally tracks, VectorDBBench and turbopuffer's harness run as published, against WarpSearch's compatible APIs.
Recall next to vector latency.A fast answer with the wrong neighbours is not a result. Where a vendor publishes recall, we compare at the same recall or higher.
Every number says what was timed.Server time and client time are labelled on every page, so you can see exactly what is being compared.
Run it on your own data.
Point your Elasticsearch, OpenSearch or turbopuffer client at WarpSearch and see the numbers for your workload. Free up to 1 GB.