Cohere-10M — 2026-06-15

r8i.8xlarge · peak recall 0.9994 @ 3 QPS

QPS at recall

Recall	Meerkat	vchordrq	vchordrq-rabitq8	vchordrq-rabitq4	Δ vs best competitor
0.90	— (max r=0.000)	225	104	42	—
0.95	— (max r=0.000)	99	57	23	—
0.99	— (max r=0.000)	21	9	8	—

Latency at recall

	Meerkat	vchordrq	vchordrq-rabitq8	vchordrq-rabitq4	Δ vs best competitor
@ recall 0.90
p50	—	4.35 ms	8.22 ms	23.33 ms	—
p95	—	5.56 ms	20.13 ms	37.11 ms	—
p99	—	6.20 ms	27.50 ms	44.79 ms	—
@ recall 0.95
p50	—	9.94 ms	14.27 ms	42.49 ms	—
p95	—	12.51 ms	37.59 ms	70.40 ms	—
p99	—	13.85 ms	51.61 ms	84.79 ms	—
@ recall 0.99
p50	—	46.70 ms	104.06 ms	124.37 ms	—
p95	—	55.51 ms	193.88 ms	232.95 ms	—
p99	—	59.42 ms	266.56 ms	277.87 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 — vchordrq @0.90: vchordrq-n16000 · vchordrq-rabitq8 @0.90: vchordrq-rabitq8-n16000 · vchordrq-rabitq4 @0.90: vchordrq-rabitq4-rerank-n16000 · vchordrq @0.95: vchordrq-n16000 · vchordrq-rabitq8 @0.95: vchordrq-rabitq8-n16000 · vchordrq-rabitq4 @0.95: vchordrq-rabitq4-rerank-n16000 · vchordrq @0.99: vchordrq-n16000 · vchordrq-rabitq8 @0.99: vchordrq-rabitq8-n16000 · vchordrq-rabitq4 @0.99: vchordrq-rabitq4-rerank-n16000

Build comparison

	Meerkat	vchordrq	vchordrq-rabitq8	vchordrq-rabitq4	Δ vs best competitor
Build time	1124 s	1216 s	1544 s	1744 s	-8%
Index size	2.3 GB	40.0 GB	9.5 GB	5.8 GB	-60%

Build stats are for the configuration backing each column's best operating point: Meerkat = mkt-n24000-cctrue-fs16-soar · vchordrq = vchordrq-n16000 · vchordrq-rabitq8 = vchordrq-rabitq8-n16000 · vchordrq-rabitq4 = vchordrq-rabitq4-rerank-n16000.

Plots

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

Host

Instance type	r8i.8xlarge
Hostname	ip-172-31-44-251
CPU	Intel(R) Xeon(R) 6975P-C (32 vCPU)
Memory	248 GB
OS	Ubuntu 26.04 LTS
Kernel	7.0.0-1006-aws

Versions

rekall	0.1.0 (27d6fe5 / 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 (cd8e895)
pg_prewarm	1.2
vchord	1.1.1
vector	0.8.2

Search

k	10
runs	3
warmup	500
num_queries	10000
S3 archive	s3://rekall-bench/runs/cohere-10m/2026-06-15/r8i.8xlarge-151840/

Index builds

Index	Type	Build time	Size	Workers
mkt-n24000-cctrue-fs16-soar	meerkat	1124.0s	2.3 GB	16
vchordrq-n16000	vchordrq	1216.0s	40.0 GB	16
vchordrq-rabitq4-rerank-n16000	vchordrq-rabitq4	1743.9s	5.8 GB	16
vchordrq-rabitq8-n16000	vchordrq-rabitq8	1543.9s	9.5 GB	16

Index parameters

mkt-n24000-cctrue-fs16-soar: {"boundary_epsilon":0.05,"centroid_compression":true,"centroid_fastscan":null,"fan_out":null,"fastscan":true,"fastscan_bits":null,"nlist":24000,"rerank_vectors":null,"soar_lambda":1.0}
vchordrq-n16000: {"build_threads":16,"lists":[16000],"rerank_n":200,"residual_quantization":true,"spherical_centroids":true}
vchordrq-rabitq4-rerank-n16000: {"build_threads":16,"lists":[16000],"rerank_n":200,"residual_quantization":false,"spherical_centroids":true}
vchordrq-rabitq8-n16000: {"build_threads":16,"lists":[16000],"rerank_n":200,"residual_quantization":false,"spherical_centroids":true}

PostgreSQL settings

effective_cache_size = 192 GB
effective_io_concurrency = 256
huge_pages = try
maintenance_work_mem = 102 GB
max_connections = 100
max_parallel_maintenance_workers = 16
max_parallel_workers = 32
random_page_cost = 1.1
shared_buffers = 64 GB
wal_level = replica
work_mem = 1 GB