Filtered-DiskANN

Overview

Filtered-DiskANN is a Microsoft Research extension to the DiskANN algorithm that enables efficient vector similarity search combined with label-based filtering, maintaining high performance even with complex filter conditions.

Key Innovation

Traditional approaches either apply filters before search (reducing candidate pool) or after search (missing relevant results). Filtered-DiskANN integrates filtering directly into the graph traversal process.

Features

• Native Filter Integration: Filters applied during graph traversal, not as pre/post-processing
• Multiple Label Support: Handle multiple filter labels per vector
• Maintained Performance: Minimal performance degradation with filters
• Guaranteed Accuracy: Ensures recall targets even with restrictive filters

Use Cases

• E-commerce search with category/price filters
• Document search with access control
• Multi-tenant vector databases
• Personalized recommendations with user constraints
• Time-based vector retrieval

Performance Characteristics

Compared to post-filtering approaches:

• Better recall at same latency
• More efficient for highly selective filters
• Reduced memory access patterns

Implementation

Implemented in:

• pgvectorscale (Timescale)
• Microsoft's DiskANN library

Research Impact

Published research has influenced modern vector database implementations, enabling practical filtered search at billion-scale.

Pricing

Research paper - free to access.