AgentDB Performance Optimization

Enable binary quantization in the AgentDB adapter configuration to achieve 32x memory reduction, turning 1M float32 vectors from 3GB into 96MB. Scalar quantization offers 4x reduction with higher accuracy, and product quantization provides 8-16x reduction for high-dimensional embeddings.

AgentDB automatically builds HNSW indices giving O(log n) search complexity, reducing search from 15ms to 100µs on 10K vectors. Tune hnswM, hnswEfConstruction, and hnswEfSearch parameters to balance recall quality against search speed for your dataset size.

Binary quantization compresses vectors 32x with 95-98% accuracy and 10x faster search, suiting memory-constrained or massive deployments. Scalar quantization compresses 4x with 98-99% accuracy and 3x faster search, fitting production apps needing high accuracy.

Yes, but small datasets benefit less from quantization. The recommended configuration for under 10K vectors uses no quantization for full precision, a 500-pattern cache, and hnswM of 8, since memory pressure is low at that scale.

Slow search usually stems from a small cache, high k values, or overly strict efSearch settings. Increase cacheSize toward 2000, reduce k from 10 to 5, or lower hnswEfSearch to 50 for faster queries at slightly reduced recall.

Avoid binary quantization when maximum accuracy matters, such as small datasets or precision-critical retrieval, since it loses 2-5% accuracy. Use scalar quantization or full precision instead, and raise hnswEfSearch to 200 for higher recall.