Just Brute Force Your Embeddings

Doug Turnbull argues that teams with moderate embedding workloads — roughly a million documents, low query traffic, and up-front embedding writes — should skip vector databases and use brute-force NumPy dot products instead.

He benchmarks 384-dimension embeddings on an M4 MacBook Pro. At one million documents, a single thread achieves ~80 queries per second with 0.012s latency; ten threads push that to ~170 QPS at 0.058s. Even at nearly nine million documents, single-thread performance remains around 9 QPS at 0.106s — adequate for many use cases.

Turnbull notes this is naive NumPy and could be faster: per-thread batching, a top-n heap for collection, or loading vectors into memory with FAISS for larger scales. His core point echoes Raymond Chen — a simple O(n) scan can beat a complex O(log n) index at practical sizes.