Vectorized MAXSCORE over WAND, especially for long LLM-generated queries (turbopuffer.com)

Turbopuffer has announced the release of FTS v2, their latest text search engine version, which boasts a remarkable performance improvement of up to 20 times faster than its predecessor, FTS v1. This significant leap is attributed mainly to an optimized search algorithm tailored for handling long queries generated by agents. In particular, FTS v2 outperforms the widely used block-max WAND algorithm in processing these complex queries, achieving several times faster execution, which is crucial as agents often input queries with many terms. The new algorithm features a variant of MAXSCORE that employs batch processing of postings iterators rather than alternating their advancement, enabling better memory locality and heightened CPU efficiency. This approach not only enhances throughput but also allows the system to leverage SIMD (Single Instruction, Multiple Data) capabilities, making it particularly effective for long queries. As the demand for efficient text search grows, especially in applications requiring extensive query handling, FTS v2 positions itself as a significant advancement in the field of AI and machine learning, optimizing search performance to keep pace with evolving user needs.