Skip to content

Quantized Centroids - #24

Merged
lkuffo merged 15 commits into
mainfrom
quantized-centroids
Oct 20, 2026
Merged

lkuffo merged 15 commits into
mainfrom
quantized-centroids

Conversation

@lkuffo

@lkuffo lkuffo commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator

Changes

  • Allows for centroids to be quantized. Starting with LVQ8, and LVQ4, which seems to be enough to maintain quality in Search and Maintenance workloads.
  • This PR tackles the main bottleneck of PDX IVF in large scale data: The distance form query to centroids.

Benchmarks

  • LVQ4 centroids ranking have 0 effect in recall.
  • LVQ4 centroids have negligible effect in maintenance.
  • LVQ4 centroids are scanned 4x faster with 8x less memory usage.

Next Steps

  • Explore RabitQ-4 as a faster option (with slightly more bookkeeping during maintenance)

@lkuffo
lkuffo merged commit 4c13ed7 into main Oct 20, 2026
15 checks passed
@lkuffo
lkuffo deleted the quantized-centroids branch October 20, 2026 16:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant