Cohere-50M — 2026-06-13

r8i.16xlarge · peak recall 0.9947 @ 16 QPS

QPS at recall

Recall	Meerkat	pgvector HNSW	pgvector HNSW (bit)	Δ vs best competitor
0.90	260	307	76	-15%
0.95	131	155	30	-15%
0.99	32	— (max r=0.988)	— (max r=0.962)	—

Latency at recall

	Meerkat	pgvector HNSW	pgvector HNSW (bit)	Δ vs best competitor
@ recall 0.90
p50	3.78 ms	3.29 ms	13.38 ms	+15%
p95	4.70 ms	4.46 ms	16.10 ms	+5%
p99	5.19 ms	5.14 ms	18.00 ms	+1%
@ recall 0.95
p50	7.49 ms	6.65 ms	34.02 ms	+13%
p95	9.29 ms	8.23 ms	40.99 ms	+13%
p99	10.16 ms	9.03 ms	43.03 ms	+12%
@ recall 0.99
p50	30.41 ms	—	—	—
p95	37.41 ms	—	—	—
p99	43.99 ms	—	—	—

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-n60000-cctrue-fs16-soar-cfs · pgvector HNSW @0.90: hnsw-m24 · pgvector HNSW (bit) @0.90: hnsw-m24-bq-rerank · Meerkat @0.95: mkt-n60000-cctrue-fs16-soar-cfs · pgvector HNSW @0.95: hnsw-m24 · pgvector HNSW (bit) @0.95: hnsw-m24-bq-rerank · Meerkat @0.99: mkt-n40000-cctrue-fs16-soar-cfs

Build comparison

	Meerkat	pgvector HNSW	pgvector HNSW (bit)	Δ vs best competitor
Build time	—	20050 s	—	—
Index size	—	190.7 GB	—	—

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

Plots

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

Host

Instance type	r8i.16xlarge
Hostname	ip-172-31-79-47
CPU	Intel(R) Xeon(R) 6975P-C (64 vCPU)
Memory	496 GB
OS	Ubuntu 26.04 LTS
Kernel	7.0.0-1006-aws

Versions

rekall	0.1.0 (44cdebe / main)
PostgreSQL	PostgreSQL 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
meerkat	0.1.0 (c9d9f4b)
pg_prewarm	1.2
vector	0.8.2
meerkat branch	fix/parallel-build-sample-gather-memory

Search

k	10
runs	3
warmup	500
num_queries	10000
S3 archive	s3://rekall-bench/runs/cohere-50m/2026-06-13/r8i.16xlarge-100020/

Index builds

Index	Type	Build time	Size	Workers
hnsw-m24	pgvector-hnsw	20050.2s	190.7 GB	32
hnsw-m24-bq	pgvector-hnsw	4869.6s	23.2 GB	32
mkt-n80000-cctrue-fs16-soar-cfs	meerkat	3547.1s	11.9 GB	32

Index parameters

hnsw-m24: {"ef_construction":200,"encoding":"vector","m":24}
hnsw-m24-bq: {"ef_construction":200,"encoding":"bit","m":24}
mkt-n80000-cctrue-fs16-soar-cfs: {"boundary_epsilon":0.05,"centroid_compression":true,"centroid_fastscan":true,"fan_out":null,"fastscan":true,"fastscan_bits":null,"nlist":80000,"rerank_vectors":null,"soar_lambda":1.0}

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