Cohere-1M — 2026-08-11

r8i.24xlarge · peak recall 1.0000 @ 67 QPS

QPS at recall

Recall	Meerkat	pgvector-ivfflat (bit)	pgvector-ivfflat (halfvec)	pgvector-ivfflat	Δ vs best competitor
0.90	1,749	102	70	48	+1615%
0.95	1,092	49	35	23	+2107%
0.99	496	12	10	7	+3923%

Latency at recall

	Meerkat	pgvector-ivfflat (bit)	pgvector-ivfflat (halfvec)	pgvector-ivfflat	Δ vs best competitor
@ recall 0.90
p50	0.55 ms	9.50 ms	13.78 ms	20.72 ms	-94%
p95	0.77 ms	12.45 ms	20.30 ms	28.76 ms	-94%
p99	0.87 ms	13.82 ms	23.19 ms	33.59 ms	-94%
@ recall 0.95
p50	0.88 ms	20.13 ms	28.07 ms	43.52 ms	-96%
p95	1.23 ms	23.38 ms	38.31 ms	56.12 ms	-95%
p99	1.41 ms	24.87 ms	42.71 ms	61.99 ms	-94%
@ recall 0.99
p50	2.00 ms	81.12 ms	102.94 ms	139.41 ms	-98%
p95	2.34 ms	86.30 ms	123.59 ms	163.19 ms	-97%
p99	2.48 ms	87.90 ms	130.07 ms	170.10 ms	-97%

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-defaults · pgvector-ivfflat (bit) @0.90: ivfflat-1m-bq-rerank · pgvector-ivfflat (halfvec) @0.90: ivfflat-1m-halfvec · pgvector-ivfflat @0.90: ivfflat-1m-fullvec · Meerkat @0.95: mkt-defaults · pgvector-ivfflat (bit) @0.95: ivfflat-1m-bq-rerank · pgvector-ivfflat (halfvec) @0.95: ivfflat-1m-halfvec · pgvector-ivfflat @0.95: ivfflat-1m-fullvec · Meerkat @0.99: mkt-defaults · pgvector-ivfflat (bit) @0.99: ivfflat-1m-bq-rerank · pgvector-ivfflat (halfvec) @0.99: ivfflat-1m-halfvec · pgvector-ivfflat @0.99: ivfflat-1m-fullvec

Build comparison

	Meerkat	pgvector-ivfflat (bit)	pgvector-ivfflat (halfvec)	pgvector-ivfflat	Δ vs best competitor
Build time	17.6 s	—	48.3 s	57.9 s	-64%
Index size	267.2 MB	—	1.5 GB	3.8 GB	-83%

Build stats are for the configuration backing each column's best operating point: Meerkat = mkt-defaults, prewarmed · pgvector-ivfflat (halfvec) = ivfflat-1m-halfvec, prewarmed · pgvector-ivfflat = ivfflat-1m-fullvec, prewarmed.

Plots

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

Host

Instance type	r8i.24xlarge
Hostname	ip-172-31-66-31
CPU	Intel(R) Xeon(R) 6975P-C (96 vCPU)
Memory	744 GB
OS	Ubuntu 26.04 LTS
Kernel	7.0.0-1008-aws

Versions

rekall	0.1.0 (aca5c4e / feature/pgvector-ivfflat)
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 (9b3e2f7)
pg_prewarm	1.2
vector	0.8.4

Search

k	10
runs	1
warmup	100
num_queries	2000
S3 archive	s3://rekall-bench/runs/cohere-1m/2026-08-11/r8i.24xlarge-143304/

Index builds

Index	Type	Build time	Size	Workers
ivfflat-1m-bq	pgvector-ivfflat	10.9s	115.8 MB	32
ivfflat-1m-fullvec	pgvector-ivfflat	57.9s	3.8 GB	32
ivfflat-1m-halfvec	pgvector-ivfflat	48.3s	1.5 GB	32
mkt-defaults	meerkat	17.6s	267.2 MB	32

Index parameters

ivfflat-1m-bq: {"encoding":"bit","lists":1000}
ivfflat-1m-fullvec: {"encoding":"vector","lists":1000}
ivfflat-1m-halfvec: {"encoding":"halfvec","lists":1000}
mkt-defaults: {"boundary_epsilon":null,"centroid_compression":null,"centroid_fastscan":null,"fan_out":null,"fastscan":null,"fastscan_bits":null,"nlist":null,"rerank_vectors":null,"soar_lambda":null}

PostgreSQL settings

effective_cache_size = 576 GB
effective_io_concurrency = 256
huge_pages = try
maintenance_work_mem = 307 GB
max_connections = 100
max_parallel_maintenance_workers = 32
max_parallel_workers = 96
random_page_cost = 1.1
shared_buffers = 192 GB
wal_level = replica
work_mem = 1 GB