Vector Indexes: Comparing IVFFlat and HNSW

Without indexes, PostgreSQL must perform a sequential scan on your table, calculating vector distance for every single row. To speed up similarity queries on large datasets, you must create vector indexes. pgvector supports two indexing algorithms: IVFFlat and HNSW.

1. IVFFlat (Inverted File Flat)

IVFFlat partitions your vector database table into clusters using k-means clustering. When you execute a search query, it checks only the most relevant clusters:

Key Characteristics

• Build Speed: Fast to generate.
• Memory Footprint: Low memory consumption.
• Accuracy vs Speed: High speed improvements but can occasionally miss nearest neighbors if the search scan parameters are set too low.
• Requirement: Requires the table to be pre-populated with data to build accurate clusters.

2. HNSW (Hierarchical Navigable Small World)

HNSW builds a multi-layered graph index where vectors are connected to their nearest neighbors. Searching this index behaves like traversing a map, quickly jumping across regions to find the closest matches:

Key Characteristics

• Query Performance: Fastest query speeds, outperforming IVFFlat by a significant margin.
• Accuracy: Excellent nearest-neighbor search accuracy.
• Memory Footprint: High memory usage because the graph structure is kept in RAM.
• Build Speed: Extremely slow index build times compared to IVFFlat.
• Requirement: Can be created on empty tables; updates dynamically as new rows are inserted.

3. Comparison Summary