Benchmark Logs
Measured, reproducible tests — existing products compared, and my own systems evaluated. Methodology shown, numbers honest. Newest first.
Which vector DB is fastest for retrieval at 1M vectors?
Setup: 1M × 768-dim vectors · top-10 retrieval · measured p95 latency over 1k queries · same hardware · default configs. Replace with your real methodology — dataset, configs, versions, what you measured, how.
| Product | p95 latency | Recall@10 | Notes |
|---|---|---|---|
| Product A | 42 ms | 0.94 | fastest here |
| Product B | 67 ms | 0.96 | best recall |
| Product C | 88 ms | 0.92 | — |
TakeawayWrite 3-5 honest lines on what the numbers showed, what surprised you, and the caveats — e.g. "this only measured latency and recall, not cost or scaling behavior; results use default configs and may differ when tuned." Honest limits are what make a benchmark credible. Replace with your own words.
Myelin retrieval — does reranking improve accuracy?
Setup: own dataset · baseline vs reranked · measured retrieval accuracy + added latency. Replace with your real methodology.
| Approach | Accuracy | Latency |
|---|---|---|
| Baseline | 0.71 | 40 ms |
| + Reranking | 0.84 | 95 ms |
TakeawayWrite your honest findings here — what the data showed about your own system, the tradeoff you found, and what you'd test next. Replace with your own words.