Use of FMA within vector similarity computations in Lucene

What is FMA

Under the hood

Is it fast

Wrapping Up

Vector Similarity Computations FMA-style

How did we introduce scalar quantization into Lucene

Automatic Byte Quantization in Lucene

Scalar Quantization 101

Nodes, Shards, Segments, oh my!

Quantization per segment

Quantization that grows with you

Performance & Numbers

Conclusion

Introducing Scalar Quantization in Lucene

What is scalar quantization and how does it work?

Buckets of fun

Time to remember your Algebra

But, how is this accurate?

Scalar quantization 101

Elasticsearch 8.11 improves how it manages its indexing buffer, resulting in less segment merging.

How it works in Elasticsearch 8.10 and earlier

Change 1: flush a single segment at a time

Change 2: flush shards in a round-robin fashion

Less merging and faster ingestion in Elasticsearch 8.11

How we brought maximum-inner-product into Lucene

Negative Values and Lucene Optimizations

The Triangle Problem

Bringing Maximum-Inner-Product into Lucene

Discover how we added passage vectors to Apache Lucene, the benefits of doing so, and how existing Lucene structures were used to create an efficient retrieval experience.

Lucene

Vector Similarity Computations FMA-style

Introducing Scalar Quantization in Lucene

Scalar quantization 101

Less merging and faster ingestion in Elasticsearch 8.11

Bringing Maximum-Inner-Product into Lucene

Adding passage vector search to Lucene