AgentDB Performance Optimization

Enable quantization when creating the AgentDB adapter. Binary quantization reduces memory 32x with 95-98% accuracy, scalar reduces 4x with 98-99% accuracy, and product quantization reduces 8-16x for high-dimensional embeddings.

AgentDB builds HNSW indices automatically, giving O(log n) search. Tune hnswM, hnswEfConstruction, and hnswEfSearch parameters: higher values improve recall but increase memory and latency, while lower values favor speed.

Use binary quantization for 1M+ vectors or memory-constrained edge deployments, scalar for balanced production workloads needing high accuracy, and product quantization for high-dimensional embeddings. Use no quantization only for small datasets requiring maximum accuracy.

Yes, each quantization type trades accuracy for compression. Binary loses roughly 2-5% accuracy, scalar loses 1-2%, and product loses 3-7%. You can recover recall by increasing hnswEfSearch at the cost of slower queries.

Slow search usually comes from a small cache, high k values, or missing quantization. Increase cacheSize toward 2000 patterns, reduce the number of results requested, and enable scalar or binary quantization to shrink the search space.

Binary quantization compresses each dimension to a single bit, so it loses 2-5% accuracy and works best on large datasets where HNSW recall compensates. It is not suitable when you need near-exact nearest neighbors on small datasets.