HNSW-IF

Overview

HNSW-IF (Hierarchical Navigable Small World - Inverted File) is Vespa's cost-efficient hybrid method for approximate nearest neighbor search, combining in-memory HNSW for centroids with disk-backed inverted file storage.

Architecture

Indexing Process

• Vectors that are not cluster centroids search for k closest centroids in HNSW graph
• Index closest centroid IDs using inverted indexes
• Inverted index dictionary maps centroid IDs to posting lists of non-centroid vector IDs

Query Process

1. Search centroid vectors using HNSW for k closest centroids
2. Search inverted index using logical disjunction (OR) of retrieved centroid IDs
3. Access disk-backed vector data via memory-mapped forward index

Cost Efficiency

• Traditional HNSW for 1B vectors (768 dims, float) requires ~4 TiB memory
• HNSW-IF keeps only centroids in memory
• Bulk of vector data stored on disk
• Dramatic cost reduction for billion-scale deployments

Performance

• Searching 200M centroid vectors with HNSW: 2-3 ms single-threaded
• Maintains high accuracy with reduced memory footprint
• Scalable to billions of vectors

Technical Details

• Vespa paged tensor attributes for disk-backed storage
• Memory-mapped forward index
• Posting lists don't contain vector data
• Efficient centroid-based retrieval

Use Cases

• Billion-scale vector search on budget
• Cost-sensitive large-scale deployments
• Filtered vector search
• Hybrid keyword + semantic search

Recognition

Noted in 2025 academic papers on vector database architectures as a leading approach for cost-efficient large-scale vector search.