← all runs

Cohere-50M — 2026-06-23

r8i.16xlarge · peak recall 0.9813 @ 73 QPS

QPS at recall

RecallMeerkatpgvector HNSWpgvector HNSW (halfvec)Δ vs best competitor
0.90— (min r=0.901)321295—
0.95193161148+20%

Latency at recall

Meerkatpgvector HNSWpgvector HNSW (halfvec)Δ vs best competitor
@ recall 0.90
p50—3.14 ms3.44 ms—
p95—4.23 ms4.54 ms—
p99—4.94 ms5.61 ms—
@ recall 0.95
p505.18 ms6.40 ms6.98 ms-19%
p956.20 ms7.99 ms8.60 ms-22%
p996.68 ms8.70 ms9.36 ms-23%

All values are taken at exactly the stated recall: each configuration's own sweep is log-interpolated between its two surrounding measured points (measured values are used verbatim on an exact hit) and the best configuration at that recall is shown — values are never interpolated across different configurations. “—” means no single sweep spans that recall; nothing is extrapolated. Recall is distance-based (ann-benchmarks, tie-aware). Insert/update/delete workloads are not measured.

Interpolated on config — pgvector HNSW @0.90: hnsw-m24 · pgvector HNSW (halfvec) @0.90: hnsw-m24-halfvec · Meerkat @0.95: mkt-n480000-cctrue-fs16-soar-cfs · pgvector HNSW @0.95: hnsw-m24 · pgvector HNSW (halfvec) @0.95: hnsw-m24-halfvec

Build comparison

Meerkatpgvector HNSWpgvector HNSW (halfvec)Δ vs best competitor
Build time—20050 s10958 s—
Index size—190.7 GB95.3 GB—

Build stats are for the configuration backing each column's best operating point: pgvector HNSW = hnsw-m24 · pgvector HNSW (halfvec) = hnsw-m24-halfvec.

Plots

Pareto svg  ·  Distance-recall svg  ·  Per-config svg  ·  Build times svg  ·  Index sizes svg
paretodistance-recallbuild-timesbuild-sizes

Host

Instance typer8i.16xlarge
Hostnameip-172-31-79-47
CPUIntel(R) Xeon(R) 6975P-C (64 vCPU)
Memory496 GB
OSUbuntu 26.04 LTS
Kernel7.0.0-1006-aws

Versions

rekall0.1.0 (44cdebe / main)
PostgreSQLPostgreSQL 18.4 (Ubuntu 18.4-1.pgdg26.04+1) on x86_64-pc-linux-gnu, compiled by gcc (Ubuntu 15.2.0-16ubuntu1) 15.2.0, 64-bit
meerkat0.1.0 (8695684)
pg_prewarm1.2
vector0.8.2

Search

k10
runs3
warmup500
num_queries3000
S3 archives3://rekall-bench/runs/cohere-50m/2026-06-23/r8i.16xlarge-130647/

Index builds

IndexTypeBuild timeSizeWorkers
hnsw-m24pgvector-hnsw20050.0s190.7 GB32
hnsw-m24-bqpgvector-hnsw4869.0s23.2 GB32
hnsw-m24-halfvecpgvector-hnsw10957.7s95.3 GB32
mkt-n80000-cctrue-fs16-soar-cfsmeerkat3540.0s11.9 GB32
Index parameters
hnsw-m24: {"ef_construction":200,"m":24}
hnsw-m24-bq: {"ef_construction":200,"encoding":"bit","m":24}
hnsw-m24-halfvec: {"ef_construction":200,"encoding":"halfvec","m":24}
mkt-n80000-cctrue-fs16-soar-cfs: {"nlist":80000}
PostgreSQL settings
effective_cache_size = 384 GB
effective_io_concurrency = 256
huge_pages = try
maintenance_work_mem = 204 GB
max_connections = 100
max_parallel_maintenance_workers = 32
max_parallel_workers = 64
random_page_cost = 1.1
shared_buffers = 128 GB
wal_level = replica
work_mem = 1 GB