How Vector Search Works

<Note>This is a beta feature available in versions 0.25.0 and above.</Note>

Today, most Postgres users rely on the pgvector extension for similarity (i.e. vector) search. pgvector works well for many use cases, but has a few limitations:

• Filtering Performance: Its indexes are separate from the ParadeDB index, so performance suffers when composing vector search with text search and other filters.
• Memory Constraints: Index build times balloon when the entire index does not fit in memory, preventing pgvector from scaling to large datasets.
• Index Quality Under Updates: Indexes can degrade over time under update-heavy workloads.

The ParadeDB index supports pgvector's vector type and is designed to solve these limitations. Vectors are indexed with a SPANN-style index, the state-of-the-art approach for billion-scale vector search.