AI News

Trending story

Modem migrates vector search from pgvector to turbopuffer

1 reports1 story updatesUpdated Nov 13 22:00

Understand this story

AI overview

Modem migrated vector search from Postgres's pgvector to turbopuffer, cutting p90 search time by 76% and offloading most of the load from the database. Previously, the HNSW index on the shared multi-tenant table would first perform a vector nearest-neighbor search across the entire table, then apply organization_id and time filters, causing small customers' queries to wait up to 10 seconds, and performance actually worsened after adding the index. By August 2026, median search time was about 0.8 seconds, with the slowest 1% reaching 28 seconds.

AI-generated from public reports

Latest development · Report

Migrating from pgvector to turbopuffer: Why the HNSW index backfired

Developments

1 confirmed developments

  1. Modem migrates vector search from pgvector to turbopuffer

    Source:Modem:Blog

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 13

  1. Modem:BlogFirst-party

    Migrating from pgvector to turbopuffer: Why the HNSW index backfired

    Modem migrated vector search from Postgres's pgvector to turbopuffer, reducing p90 search time by 76% and offloading most of the load from the database. Previously, an HNSW index on a shared multi-tenant table would first perform a vector nearest-neighbor search across the entire table and then apply organization_id and time filters, causing queries from small customers to wait up to 10 seconds, and performance actually got worse after adding the index. By August 2026, median search time was about 0.8 seconds, with the slowest 1% reaching 28 seconds.

Related stories

  1. Related storiesDatabricks showcases unified platform solution for real-time energy trading and analytics
  2. Related storiesEnglish translation is being prepared.
  3. Related storiesIndegene builds an LLM data quality Agent in NEXT UDP
  4. Related storiesMiles summarizes the eight-step multimodal data cleaning solution for FDE at beauty companies