← all runs
Cohere-50M — 2026-06-26
r8i.12xlarge · peak recall 0.9930 @ 27 QPS
QPS at recall
| Recall | Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| 0.90 | 253 | 276 | — (min r=0.953) | -8% |
| 0.95 | 153 | 146 | — (min r=0.953) | +5% |
| 0.99 | 35 | — (max r=0.985) | — (max r=0.956) | — |
Latency at recall
| Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| @ recall 0.90 | | | | |
| p50 | 3.83 ms | 3.67 ms | — | +5% |
| p95 | 5.65 ms | 4.81 ms | — | +17% |
| p99 | 7.01 ms | 5.56 ms | — | +26% |
| @ recall 0.95 | | | | |
| p50 | 6.35 ms | 7.11 ms | — | -11% |
| p95 | 8.86 ms | 8.74 ms | — | +1% |
| p99 | 10.34 ms | 9.86 ms | — | +5% |
| @ recall 0.99 | | | | |
| p50 | 27.49 ms | — | — | — |
| p95 | 36.58 ms | — | — | — |
| p99 | 43.10 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-n120000-cctrue-fs16-soar-cfs · pgvector HNSW (halfvec) @0.90: hnsw-m24-halfvec · Meerkat @0.95: mkt-n240000-cctrue-fs16-soar-cfs · pgvector HNSW (halfvec) @0.95: hnsw-m24-halfvec · Meerkat @0.99: mkt-n240000-cctrue-fs16-soar-cfs
Build comparison
| Meerkat | pgvector HNSW (halfvec) | pgvector HNSW (bit) | Δ vs best competitor |
|---|
| Build time | 2738 s | 13372 s | — | -80% |
| Index size | 12.8 GB | 95.3 GB | — | -87% |
Build stats are for the configuration backing each column's best operating point: Meerkat = mkt-n240000-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.12xlarge |
| Hostname | ip-172-31-79-49 |
| CPU | Intel(R) Xeon(R) 6975P-C (48 vCPU) |
| Memory | 372 GB |
| OS | Ubuntu 26.04 LTS |
| Kernel | 7.0.0-1006-aws |
Versions
| rekall | 0.1.0 (ce00e5f / 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 (d5d6040) |
| pg_prewarm | 1.2 |
| vector | 0.8.3 |
Search
Index builds
| Index | Type | Build time | Size | Workers |
|---|
| hnsw-m24-bq | pgvector-hnsw | 5430.0s | 23.2 GB | 24 |
| hnsw-m24-halfvec | pgvector-hnsw | 13372.0s | 95.3 GB | 24 |
| mkt-n120000-cctrue-fs16-soar-cfs | meerkat | 6795.9s | 11.9 GB | 24 |
| mkt-n240000-cctrue-fs16-soar-cfs | meerkat | 2738.3s | 12.8 GB | 24 |
Index parameters
hnsw-m24-bq: {"ef_construction":200,"encoding":"bit","m":24}
hnsw-m24-halfvec: {"ef_construction":200,"encoding":"halfvec","m":24}
mkt-n120000-cctrue-fs16-soar-cfs: {"boundary_epsilon":0.05,"centroid_compression":true,"centroid_fastscan":true,"fan_out":null,"fastscan":true,"fastscan_bits":null,"nlist":120000,"rerank_vectors":null,"soar_lambda":1.0}
mkt-n240000-cctrue-fs16-soar-cfs: {"boundary_epsilon":0.05,"centroid_compression":true,"centroid_fastscan":true,"fan_out":null,"fastscan":true,"fastscan_bits":null,"nlist":240000,"rerank_vectors":null,"soar_lambda":1.0}
PostgreSQL settings
effective_cache_size = 288 GB
effective_io_concurrency = 256
huge_pages = try
maintenance_work_mem = 153 GB
max_connections = 100
max_parallel_maintenance_workers = 24
max_parallel_workers = 48
random_page_cost = 1.1
shared_buffers = 96 GB
wal_level = replica
work_mem = 1 GB
