Customization

Cosdata offers extensive customization options to tailor the system to your specific requirements.

Vector Configurations

Dimension Settings

Cosdata supports vectors of various dimensions:

• Low-dimensional (32-256): Ideal for simple embeddings
• Medium-dimensional (256-768): Suitable for most language models
• High-dimensional (768+): For advanced embedding models

When creating a collection, you can specify the exact dimension:

{
    "name": "testdb",
    "description": "Test collection for vector database",
    "dense_vector": {
        "enabled": true,
        "auto_create_index": false,
        "dimension": 1024
    },
    "sparse_vector": {
        "enabled": false,
        "auto_create_index": false
    }
}

Distance Metrics

Choose from multiple distance metrics:

• Euclidean (L2)
• Cosine Similarity
• Dot Product
• Manhattan (L1)

When creating an index, you can specify the distance metric:

{
    "collection_name": "testdb",
    "name": "testdb_index",
    "distance_metric_type": "cosine",
    "quantization": "scalar",
    "data_type": "u8",
    "index_type": "hnsw"
}

Index Tuning

HNSW Parameters

Fine-tune your HNSW indexes with these advanced parameters:

{
    "collection_name": "testdb",
    "name": "testdb_index",
    "distance_metric_type": "cosine",
    "quantization": "scalar",
    "data_type": "u8",
    "index_type": "hnsw",
    "params": {
        "num_layers": 5,
        "max_cache_size": 1000
    }
}

Key parameters include:

• num_layers: Controls the number of layers in the HNSW graph
• max_cache_size: Maximum number of vectors to cache for performance

Filtering Options

Implement custom filters to narrow search results:

# Example: Searching with metadata filters
results = client.search(
    collection_name="products",
    query_vector=embedding,
    filter={
        "category": "electronics",
        "price": {"$lt": 1000}
    },
    limit=10
)

Transaction Management

Cosdata provides ACID transaction guarantees for vector operations, allowing you to:

1. Group related operations in a single transaction
2. Ensure atomicity across multiple vector operations
3. Maintain consistency during concurrent access

Transaction Best Practices

For optimal performance:

1. Keep transaction duration short
2. Batch vector operations (100-1000 vectors per batch)
3. Always commit or abort transactions to release resources
4. Implement proper error handling with retry logic

Deployment Options

Cosdata can be deployed in various configurations:

• Standalone: Single-node deployment for simplicity
• Distributed: Multi-node deployment for high availability
• Hybrid: Combination of on-premise and cloud resources

Performance Optimization

For production deployments, consider these optimization strategies:

1. Use parallel requests for large datasets
2. Monitor response times and transaction timeouts
3. Index important vector fields
4. Normalize vectors to unit length (values between -1.0 and 1.0)
5. Choose appropriate similarity metrics for your use case