Benchmarks

Test Environment

Hardware: Intel i9-13900K, 64GB RAM
Backend: usearch HNSW v2.12+
Vectors: 384 dimensions (Snowflake/snowflake-arctic-embed-xs)

Search Speed (k=10)

Comparison of usearch HNSW approximate search vs brute-force exact search:

Collection Size	usearch (ms)	Brute-force (ms)	Speedup	Recall@10
1,000	0.11	0.11	1.0x	1.00
5,000	0.33	0.34	1.0x	1.00
10,000	0.23	0.82	3.6x	1.00
50,000	0.26	4.67	18x	0.70
100,000	0.27	9.18	34x	0.60

Note: Collections <10k use adaptive brute-force for perfect recall by default.

Storage Efficiency (Quantization)

10,000 vectors at 384 dimensions:

Quantization	Total Size	Query Latency	Vector Compression	Notes
FLOAT32	36.0 MB	0.20 ms	1x (baseline)	Full precision
FLOAT16	28.7 MB	0.20 ms	2x	Best balance for most use cases
INT8	25.0 MB	0.16 ms	4x	Good for memory-constrained
BIT	21.8 MB	0.08 ms	32x	Fastest, Hamming distance

Note: Total size includes SQLite metadata + HNSW graph structure (~21MB overhead). Pure vector storage follows theoretical compression ratios.

Adaptive Search Behavior

SimpleVecDB v2.0 automatically selects the optimal search strategy:

Collection Size	Default Strategy	Rationale
< 10,000	Brute-force	Perfect recall, HNSW overhead not worth it
≥ 10,000	HNSW	3-34x faster, tunable recall

Override with the exact parameter:

# Force brute-force (perfect recall)
results = collection.similarity_search(query, k=10, exact=True)

# Force HNSW (faster)
results = collection.similarity_search(query, k=10, exact=False)

Batch Search Performance

Use similarity_search_batch() for multi-query workloads:

Batch Size	Sequential Time	Batch Time	Speedup
10	2.0 ms	0.5 ms	4x
100	20.0 ms	3.0 ms	7x
1000	200.0 ms	25.0 ms	8x

queries = [query1, query2, query3]  # List of embedding vectors
batch_results = collection.similarity_search_batch(queries, k=10)

Memory-Mapping Behavior

Large indexes automatically use memory-mapped mode for instant startup:

Collection Size	Mode	Startup Time	Memory Usage
< 100,000	Load	~50ms	Full index
≥ 100,000	Mmap	~1ms	On-demand

Quantization Guide

Quantization	Bits/Dim	Memory	Speed	Recall	Best For
FLOAT32	32	1x	1x	100%	Maximum precision
FLOAT16	16	0.5x	1x	~99%	Balanced (recommended)
INT8	8	0.25x	1.2x	~97%	Memory-constrained
BIT	1	0.03x	2.5x	~85%	Massive scale, fast filtering

Legacy Benchmarks (v1.x with sqlite-vec)

For historical reference, here are the v1.x benchmarks using sqlite-vec brute-force search.

10,000 Vectors (sqlite-vec v0.1.6)

Quantization	Storage	Insert Speed	Query (k=10)
FLOAT32	15.5 MB	15,585 vec/s	3.55 ms
INT8	4.2 MB	27,893 vec/s	3.93 ms
BIT	0.95 MB	32,321 vec/s	0.27 ms

100,000 Vectors (sqlite-vec v0.1.6)

Quantization	Storage	Insert Speed	Query (k=10)
FLOAT32	151.8 MB	9,513 vec/s	38.73 ms
INT8	41.4 MB	13,213 vec/s	39.08 ms
BIT	9.3 MB	14,334 vec/s	1.96 ms

Running Your Own Benchmarks

# Install with server extras for embedding generation
pip install "simplevecdb[server]"

# Run the backend benchmark
python examples/backend_benchmark.py

# Run quantization benchmark  
python examples/quant_benchmark.py

# Run full performance benchmark
python examples/embeddings/perf_benchmark.py