Vector Databases Benchmarked on HNSW Indexes

Milvus, Qdrant, Pinecone, and PGVector Evaluated Across Recall@10, QPS, and Memory Footprint

Empirical Architecture Comparison: Vector Database Benchmark: 1 Million 1536-Dimensional Vectors

1. Approximate Nearest Neighbor Search & The Curse of Dimensionality

In embedding-based AI applications (such as OpenAI text-embedding-3 or Cohere Embed), each document chunk is mapped to a vector in $\mathbb{R}^{1536}$. Calculating cosine distance across 1,000,000 documents requires 1.5 billion floating-point operations per query: $$\text{Cosine Similarity} = \frac{\mathbf{u} \cdot \mathbf{v}}{\|\mathbf{u}\| \|\mathbf{v}\|}$$ At scale, brute-force linear scanning induces massive CPU contention and 150ms query latencies. Approximate Nearest Neighbor (ANN) search trades a fraction of percentage point recall for 100x query speedups.

2. The Mechanics of Hierarchical Navigable Small World (HNSW) Graphs

HNSW builds a multi-layer graph inspired by Skip Lists. The top layers contain sparse links spanning long geometric distances, allowing logarithmic-time global navigation. The bottom layer ($L_0$) contains dense local connections:

1. A query enters at the top layer and performs greedy routing, jumping to the neighbor closest to the query vector.
2. When no closer neighbor exists on the current layer, routing steps down to the layer below.
3. At layer 0, a local search collects the top-$k$ nearest neighbors.

Key parameters control the precision-speed trade-off: $M$ (maximum connections per node, typically 16–64) and `efConstruction` (size of the dynamic candidate list during build time, typically 100–200).

3. Inverted File Quantization (IVF-PQ) vs. HNSW

While HNSW delivers the highest QPS and Recall, it requires storing vector graphs in RAM, consuming up to 8GB of memory per million vectors. In memory-constrained environments, Product Quantization (PQ) decomposes 1536-dimensional vectors into 96 sub-vectors of 16 dimensions each, quantizing each sub-vector into an 8-bit centroid index: $$\text{Compression Ratio} = \frac{1536 \times 32 \text{ bits}}{96 \times 8 \text{ bits}} = 64\times \text{ memory reduction}$$ This slashes RAM consumption from 8GB down to 125MB, with an acceptable 4% drop in Recall@10.

4. Architectural Recommendation for Production Deployments

For teams already running PostgreSQL in production with sub-million vector datasets, `pgvector` with HNSW indexes eliminates infrastructure sprawl and simplifies ACID transactions. For high-throughput enterprise applications (>5M vectors, >2,000 QPS), dedicated engines like Qdrant (Rust) or Milvus (C++) provide the necessary horizontal sharding, scalar filtering optimizations, and sub-3ms latency guarantees.