Skip to content
Cosdata Documentation

Benchmarks

Cosdata delivers exceptional performance compared to other vector databases. The following benchmarks demonstrate Cosdata’s capabilities in real-world scenarios.

Benchmark Comparison

These benchmarks were conducted using the Qdrant benchmark suite, which provides standardized tests for vector database performance.

Key Performance Metrics

MetricCosdataNext Best CompetitorImprovement
Requests Per Second (RPS)17731238 (Qdrant)43% faster
Average Recall@598.08%Similar (0.97-0.99)Comparable accuracy
Upload + Index Time14.98 minutes13.04 minutes (Weaviate)Competitive
Precision0.980.97-0.99Industry standard

The complete benchmark data comparing all tested vector databases is available on the Qdrant benchmarks page.

Additional Cosdata Performance Tests

We’ve conducted additional benchmarks on different hardware configurations and with various parameter settings to demonstrate Cosdata’s versatility:

Million Embeddings Test (768 dimensions)

Testing environment:

  • Dataset: 1 million text embeddings (768 dimensions)
  • Hardware: x86 machine, 4C/8T, 32 GB RAM
ConfigurationUpload + Index TimeAverage Recall@5RPS
High Recall
ef_construction: 128, ef_search: 128
neighbors: 16, layer_0_neighbors: 32		9.23 minutes97.60%2274.31
Balanced
ef_construction: 64, ef_search: 128
neighbors: 16, layer_0_neighbors: 32		8.64 minutes98.20%2242.33
Fast Search
ef_construction: 64, ef_search: 64
neighbors: 16, layer_0_neighbors: 32		7.80 minutes95.60%2621.02
Maximum Speed
ef_construction: 32, ef_search: 32
neighbors: 16, layer_0_neighbors: 32		7.05 minutes94.80%2959.85

Performance Insights

These additional tests reveal important characteristics of Cosdata:

  1. Configurable Performance Tradeoffs: By adjusting parameters like ef_construction and ef_search, you can optimize for either higher recall accuracy or faster indexing and search.

  2. Hardware Scaling: Performance scales well with hardware resources. The 4-core machine achieved nearly 3000 RPS with optimized settings.

  3. Dimension Handling: Cosdata efficiently handles both medium (768) and high (1536) dimensional vectors with excellent performance.

What These Results Mean

These benchmarks demonstrate Cosdata’s exceptional performance characteristics:

  1. High Throughput: With up to 2959 RPS on modest hardware, Cosdata outperforms other vector databases, handling significantly more queries per second than competitors.

  2. Excellent Recall with High Precision: Cosdata achieves 98%+ recall with high precision across configurations, providing accurate search results without sacrificing performance.

  3. Efficient Indexing: Cosdata’s indexing performance balances speed and quality, creating efficient indexes in minutes rather than hours.

  4. Flexible Configuration: The ability to tune parameters allows you to optimize for your specific use case, whether you need maximum accuracy or maximum throughput.

Benchmark Methodology

The benchmarks were conducted using:

  • Datasets:
    • dbpedia-openai-1M-1536-angular (1 million 1536-dimensional OpenAI embeddings)
    • million-text-embeddings (1 million 768-dimensional text embeddings)
  • Hardware: Standardized cloud instances and on-premise hardware with consistent specifications
  • Metrics:
    • Upload time: Time to upload vectors
    • Index time: Time to build search indexes
    • Recall@5: Percentage of correct results in top 5 results
    • RPS: Requests per second (throughput)
    • Precision: Accuracy of results compared to ground truth

Learn More