forge-db

Features

• IVF-PQ indexing with residual quantization
• SIMD-accelerated distance functions (AVX2/AVX-512 with runtime detection)
• 16x memory compression via Product Quantization (8-bit) or 32x with 4-bit PQ
• Optional re-ranking for higher recall
• Batch query optimization with partition-centric processing
• Lock-free HNSW search after index finalization
• Profile-guided optimization (PGO) build support
• Pure Rust, no external dependencies for core algorithm

Benchmarks

SIFT1M dataset (1M vectors, 128 dimensions):

Mode	Latency	QPS	Recall@10	Memory
nprobe=1	75 µs	13,333	33%	31 MB
nprobe=16	520 µs	1,923	69%	31 MB
nprobe=16 + rerank	560 µs	1,786	91%	31 MB
nprobe=32 + rerank	1.0 ms	1,000	96%	31 MB

Uncompressed memory: 488 MB. Compressed: 31 MB (16x reduction).

SIMD distance improvements (128 dimensions): 3.2x faster vs scalar.

Usage

use forge_db::{Vector, IVFPQIndex, DistanceMetric};

// Build index
let vectors: Vec<Vector> = load_vectors();
let mut index = IVFPQIndex::build(
    vectors.clone(),
    1024,  // clusters
    32,    // subvectors (compression)
    DistanceMetric::Euclidean,
);

// Optional: enable re-ranking for higher recall
index.enable_reranking(vectors, 4);

// Search
index.set_nprobe(16);
let results = index.search(&query, 10);

Build

cargo build --release
cargo test
cargo bench --bench pq_bench

For maximum performance, use profile-guided optimization:

./pgo-build.sh

Run SIFT1M benchmark

# Download dataset
curl -O ftp://ftp.irisa.fr/local/texmex/corpus/sift.tar.gz
tar -xzf sift.tar.gz
mkdir -p data && mv sift data/sift1m

# Run benchmark
cargo run --release --example sift1m_benchmark

License

MIT