Skip to content

vector-search.md

Latest commit

 

History

History
90 lines (66 loc) · 4.09 KB

vector-search.md

File metadata and controls

90 lines (66 loc) · 4.09 KB
Raw

Vector Search Benchmark

CLI name: vector_search - use benchbox run --benchmark vector_search

Query templates: every Vector Search query

The Vector Search benchmark tests similarity-search performance across OLAP databases that support array/vector operations. It covers exact kNN, filtered kNN, and approximate (ANN) search over synthetic embedding data.

Overview

Property Value
Total Queries 6
Data Source Synthetic embeddings (reproducible)
Default Dimensions 128
Primary Target DuckDB (array_cosine_similarity, array_distance)
Dialect Variants pgvector, ClickHouse, Snowflake, StarRocks, Doris

Query Coverage

ID / Area Description
Exact kNN (cosine) Exact k-nearest neighbors by cosine similarity
Exact kNN (L2) Exact k-nearest neighbors by Euclidean distance
Filtered kNN Metadata predicate + distance ordering
Large-k ground truth Ground-truth generation for recall evaluation
ANN search Same SQL form; relies on HNSW index from load phase
Multi-category filtered Multi-predicate filtered similarity search

Scale Factors

Scale Vectors Notes
0.01 ~10 K Quick smoke test
0.1 ~100 K Dev / CI
1.0 ~1 M Default - representative workload

Usage

benchbox run --platform duckdb --benchmark vector_search --scale 1.0

benchbox run --platform duckdb --benchmark vector_search --scale 0.1 \
  --benchmark-option dimensions=64

benchbox run --platform duckdb --benchmark vector_search --scale 1.0 \
  --benchmark-option dimensions=768

The first command uses the defaults: 128-dimensional vectors at SF=1 (about 1M vectors). The second uses smaller vectors for quick tests, and the third uses higher-dimensional embeddings.

Benchmark Options

Option Default Description
dimensions 128 Embedding vector dimensionality

Platform Support

Vector Search targets engines with native array / vector operations:

Platform Support Notes
DuckDB Primary Native array_cosine_similarity / array_distance
PostgreSQL Via pgvector Requires pgvector extension
ClickHouse Yes Dialect variant uses ClickHouse vector ops
Snowflake Yes Dialect variant uses Snowflake vector ops
StarRocks Yes Dialect variant uses StarRocks vector distance fns
Doris Yes Dialect variant uses Doris array distance fns

Notes

  • Synthetic data is reproducible - the same seed produces identical embeddings across runs.
  • ANN queries reuse kNN SQL but rely on an HNSW (or equivalent) index being built during the load phase.
  • Recall is evaluated against the large-k ground-truth query; compare approximate vs. exact results after a run.

See Also