← all runs

Cohere-10M — 2026-06-12

r8i.2xlarge · peak recall 0.9962 @ 16 QPS

QPS at recall

RecallMeerkatpgvector HNSWpgvector HNSW (bit)Δ vs best competitor
0.90579352146+64%
0.9530718881+63%
0.998643— (max r=0.981)+99%

Latency at recall

Meerkatpgvector HNSWpgvector HNSW (bit)Δ vs best competitor
@ recall 0.90
p501.71 ms2.80 ms6.91 ms-39%
p952.11 ms4.01 ms8.58 ms-47%
p992.30 ms4.90 ms9.48 ms-53%
@ recall 0.95
p503.25 ms5.36 ms12.54 ms-39%
p953.86 ms6.89 ms15.19 ms-44%
p994.17 ms7.86 ms16.02 ms-47%
@ recall 0.99
p5011.74 ms23.66 ms—-50%
p9513.29 ms28.89 ms—-54%
p9913.81 ms30.76 ms—-55%

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 — Meerkat @0.90: mkt-n24000-cctrue-fs16-soar-cfs · pgvector HNSW @0.90: hnsw-m24 · pgvector HNSW (bit) @0.90: hnsw-m24-bq-rerank · Meerkat @0.95: mkt-n24000-cctrue-fs16-soar-cfs · pgvector HNSW @0.95: hnsw-m24 · pgvector HNSW (bit) @0.95: hnsw-m24-bq-rerank · Meerkat @0.99: mkt-n24000-cctrue-fs16-soar-cfs · pgvector HNSW @0.99: hnsw-m24

Plots

Pareto svg  ·  Distance-recall svg  ·  Per-config svg
paretodistance-recall

Host

Instance typer8i.2xlarge
Hostnameip-172-31-71-30
CPUIntel(R) Xeon(R) 6975P-C (8 vCPU)
Memory62 GB
OSUbuntu 26.04 LTS
Kernel7.0.0-1006-aws

Versions

rekall0.1.0 (11faabd / 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 (8e85a2b)
pg_prewarm1.2
vector0.8.2

Search

k10
runs3
warmup500
num_queries10000
S3 archives3://rekall-bench/runs/cohere-10m/2026-06-12/r8i.2xlarge-214404/
PostgreSQL settings
effective_cache_size = 48 GB
effective_io_concurrency = 256
huge_pages = try
maintenance_work_mem = 25 GB
max_connections = 100
max_parallel_maintenance_workers = 4
max_parallel_workers = 8
random_page_cost = 1.1
shared_buffers = 16 GB
wal_level = replica
work_mem = 512 MB