Performance

Cosdata is designed from the ground up for exceptional performance, even at scale. Our benchmarks demonstrate industry-leading throughput and efficiency.

Benchmark Results

Dense Vector Search

Cosdata sets the standard for dense vector search performance:

Metric	Value	Comparison
QPS	1,758+	42% faster than Qdrant, 54% faster than Weaviate, 146% faster than ElasticSearch
Precision	97%	Consistent across challenging search tasks
Index Creation	Up to 12x faster	vs. ElasticSearch on large datasets
Latency	Lower p50/p95	Across all tested datasets

Cosdata delivers industry-leading throughput and accuracy for dense vector workloads.

Full-Text Search (BM25)

Cosdata’s custom BM25 implementation delivers:

Metric	Value	Comparison
QPS	Up to 151x faster	vs. ElasticSearch (scifact dataset), ~44x average across all datasets
Ranking Quality (NDCG)	Similar	To ElasticSearch
Index Creation	Up to 12x faster	vs. ElasticSearch on large datasets
Latency	Lower p50/p95	Across all tested datasets

Cosdata matches or exceeds ElasticSearch in ranking quality while providing dramatically higher throughput and lower latency.

For complete benchmark details, see our Benchmarks page (opens in a new tab).

Optimization Techniques

HNSW Indexing

Cosdata uses Hierarchical Navigable Small World (HNSW) graphs for approximate nearest neighbor search, providing:

• Logarithmic search complexity
• Configurable precision/speed tradeoff
• Efficient memory usage

Parallel Processing

Cosdata leverages multi-threading and SIMD instructions to maximize performance:

• Parallel query processing
• Vectorized distance calculations
• Efficient memory access patterns

Smart Quantization

Reduce memory usage while maintaining search quality:

• Scalar quantization for moderate compression
• Product quantization for higher compression needs
• Configurable precision/memory tradeoffs

Scaling Guidelines

For optimal performance as your data grows:

1. Hardware Recommendations:

   • 16+ CPU cores
   • 64GB+ RAM
   • NVMe SSD storage

2. Configuration Tuning:

   • Adjust HNSW parameters (M, efConstruction, efSearch)
   • Optimize batch sizes for your workload
   • Configure memory limits appropriately

Performance Monitoring

Cosdata provides built-in performance monitoring tools:

# Get performance metrics
curl http://localhost:8443/metrics

Key metrics to monitor:

• Query latency (p95, p99)
• Throughput (requests per second)
• Memory usage
• Index size

Performance Optimization Tips

Batch Operations

Use batch operations for better throughput:

# Batch insertion is more efficient than individual inserts
client.batch_insert_vectors(
    collection_name="my_collection",
    vectors=batch_vectors  # List of vectors
)

Index Tuning

Optimize your indexes for your specific workload:

# For higher recall at the cost of indexing time
client.create_index(
    collection_name="my_collection",
    index_name="high_recall_index",
    index_type="hnsw",
    parameters={
        "M": 32,              # Higher M = more connections = better recall
        "efConstruction": 200, # Higher efConstruction = better index quality
        "efSearch": 100,      # Higher efSearch = better search quality
        "metric": "cosine"
    }
)

# For faster indexing and lower memory usage
client.create_index(
    collection_name="my_collection",
    index_name="fast_index",
    index_type="hnsw",
    parameters={
        "M": 16,              # Lower M = fewer connections = faster indexing
        "efConstruction": 100, # Lower efConstruction = faster indexing
        "efSearch": 50,       # Lower efSearch = faster search
        "metric": "cosine"
    }
)

Vector Normalization

Normalize vectors before insertion for better search quality:

import numpy as np

def normalize_vector(vector):
    """Normalize a vector to unit length."""
    norm = np.linalg.norm(vector)
    if norm == 0:
        return vector
    return vector / norm

# Normalize before insertion
normalized_vector = normalize_vector(my_vector)
client.insert_vector(
    collection_name="my_collection",
    vector_id="doc1",
    vector=normalized_vector.tolist()
)