250 Million Benchmark

[brian-ogrady]

Wikipedia cohere 1024 dimensional dataset

• Ingest 250M records
• Conduct load test on full-resolution vectors
  • QPS
  • p99 latency
  • recall @ k
• Conduct same load test with following quantization options
  • Binary Quantization
  • Scalar Quantization
  • Binary two-bit
  • Binary 1.5-bit
  • TurboQuant (new feature)
• Run same benchmarks with varying m and ef_construct parameters