Skip to content

performance.md

Latest commit

 

History

History
71 lines (54 loc) · 2.36 KB

performance.md

File metadata and controls

71 lines (54 loc) · 2.36 KB
Raw

Performance Evaluation

Methodology

  • All benchmarks run via Criterion.rs with statistical analysis
  • Hardware: [to be filled by runner]
  • Rust version: 1.75+ with release profile

Microbenchmarks

Operation Target Measured Status
Single embed (384D) < 5ms TBD
Batch embed (100) < 50ms TBD
Vector search (1K) < 1ms TBD
Vector search (10K) < 5ms TBD
Vector search (100K) < 10ms TBD
Cosine similarity (384D) < 1us TBD
Categorize (10 centroids) < 1ms TBD
Scalar quantize roundtrip < 1us TBD

Search Scaling

Index Size p50 Target p95 Target p99 Target Measured
1K vectors < 1ms < 2ms < 5ms TBD
10K vectors < 5ms < 10ms < 20ms TBD
100K vectors < 10ms < 25ms < 50ms TBD

Ingestion Pipeline

Mode Target Measured
Text-only > 500 emails/sec TBD
Multi-asset > 50 emails/sec TBD

Memory Profile

Corpus Size Target Measured
10K emails ~115MB TBD
100K emails ~700MB TBD

Quantization Impact

Tier Compression Recall@10 Search Speed
None (fp32) 1x baseline TBD
Scalar (int8) 4x ~99.5% TBD
Binary (1-bit) 32x ~90% TBD

Clustering Scaling

Data Size Target Measured
100 vectors < 10ms TBD
1K vectors < 100ms TBD
10K vectors < 2s TBD

How to Run

# Run all benchmarks
cargo bench --bench vector_benchmarks

# Run a specific benchmark group
cargo bench --bench vector_benchmarks -- search_scaling
cargo bench --bench vector_benchmarks -- ingestion_pipeline
cargo bench --bench vector_benchmarks -- quantization_comparison
cargo bench --bench vector_benchmarks -- clustering_scaling