Real-world benchmark suites for QubicDB — measuring retrieval quality, latency, concurrency, and hybrid search accuracy.

All results are reproducible via Docker Compose. No cloud dependency.

End-to-end retrieval benchmark using a biology knowledge dataset. Measures write throughput, search latency percentiles, and Top-K recall against a domain-specific corpus.

Run:

cd biology && bash run.sh

Compares lexical-only vs hybrid (vector + lexical) search across two independent indexes. Measures Top-1 accuracy, Top-5 hit rate, MRR, Precision@5, latency percentiles, and cross-index consistency.

Latest results (2026-02-19 — 500 iterations, 750 queries/index):

Hybrid search matches lexical accuracy while delivering significantly lower latency and better cross-index consistency.

Run:

cd deepage
bash run-lexical-prefix.sh --iterations=500
bash run-vector-prefix.sh  --iterations=500

See deepage/results/conclusion.md for full analysis.

Engine and concurrency microbenchmarks (no Docker, no model required):

# Engine microbenchmarks
go test -run '^$' -bench 'BenchmarkMatrixEngine' \
  -benchmem ./pkg/engine

# Worker/concurrency path
go test -run '^$' -bench 'BenchmarkBrainWorker' \
  -benchmem ./pkg/concurrency

Results (Apple M3 arm64):

• Docker + Docker Compose (for biology/ and deepage/)
• Go 1.22+ (for microbenchmarks)
• GGUF model file at qubicdb/dist/MiniLM-L6-v2.Q8_0.gguf (for vector benchmarks only)

MIT

Developed by Deniz Umut Dereli