Vector Indexes

Vector indexes are specialized data structures that enable efficient similarity search across your content embeddings. They are optimized for different modalities and vector types in our model registry.

​

Understanding Vector Indexes

​

Key Concepts

• Vector Space: Multidimensional space where content is represented, with dimensions varying by model (512-1408)
• Modality Support: Different indexes optimized for specific content types (text, image, video, audio)
• Vector Types: Support for both dense and sparse embeddings
• Performance Tradeoffs: Balance between speed and accuracy

​

Available Index Types

​

1. HNSW (Hierarchical Navigable Small World)

Best for:

• Dense vectors from our image (512d), text (1024d), and multimodal (1408d) models
• Real-time search applications
• Large-scale datasets

​

2. SPLADE (Sparse Lexical AnD Expansion)

Best for:

• Sparse vectors from our keyword model
• Text-specific search
• Interpretable results

​

3. Flat Index

Best for:

• Small datasets
• Exact search requirements
• Development/testing
• Maximum accuracy needed