← all runs
Cohere-100M — 2026-07-07
r8i.24xlarge · peak recall 0.9872 @ 59 QPS
QPS at recall
| Recall | Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| 0.90 | 427 | 270 | — (min r=0.947) | +58% |
| 0.95 | 207 | 130 | — (max r=0.950) | +60% |
Latency at recall
| Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| @ recall 0.90 | | | | |
| p50 | 2.32 ms | 3.74 ms | — | -38% |
| p95 | 2.88 ms | 5.00 ms | — | -42% |
| p99 | 3.16 ms | 6.04 ms | — | -48% |
| @ recall 0.95 | | | | |
| p50 | 4.77 ms | 8.00 ms | — | -40% |
| p95 | 5.90 ms | 9.83 ms | — | -40% |
| p99 | 6.46 ms | 10.87 ms | — | -41% |
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-n400000-cctrue-fs16-soar-cfs · pgvector HNSW (halfvec) @0.90: hnsw-m24-halfvec · Meerkat @0.95: mkt-n400000-cctrue-fs16-soar-cfs · pgvector HNSW (halfvec) @0.95: hnsw-m24-halfvec
Build comparison
| Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| Build time | 37917 s | 22012 s | — | +72% |
| Index size | 20.9 GB | 179.2 GB | — | -88% |
Build stats are for the configuration backing each column's best operating point: Meerkat = mkt-n400000-cctrue-fs16-soar-cfs · pgvector HNSW (halfvec) = hnsw-m24-halfvec.
Plots
Pareto svg ·
Distance-recall svg ·
Per-config svg ·
Build times svg ·
Index sizes svgHost
| 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 (8aa6a65 / 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 (4a39b1c) |
| pg_prewarm | 1.2 |
| vector | 0.8.4 |
Search
Index builds
| Index | Type | Build time | Size | Workers |
|---|
| hnsw-m24-bq | pgvector-hnsw | 8416.0s | 43.5 GB | 32 |
| hnsw-m24-halfvec | pgvector-hnsw | 22012.2s | 179.2 GB | 32 |
| mkt-n400000-cctrue-fs16-soar-cfs | meerkat | 37916.6s | 20.9 GB | 32 |
Index parameters
hnsw-m24-bq: {"ef_construction":200,"encoding":"bit","m":24}
hnsw-m24-halfvec: {"ef_construction":200,"encoding":"halfvec","m":24}
mkt-n400000-cctrue-fs16-soar-cfs: {"boundary_epsilon":0.05,"centroid_compression":true,"centroid_fastscan":true,"fan_out":null,"fastscan":true,"fastscan_bits":null,"nlist":400000,"rerank_vectors":null,"soar_lambda":1.0}
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
