Skip to content- Sources: primary, discussion
- Summary: The post names what keying every document by its ANN address costs: multi-vector documents duplicate their contents per vector, SPFresh rebalancing cascades into moving full documents and their inverted indexes, and every query plan inherits the ANN index's 100 to 200 document block size. The load-bearing evidence is already shipped rather than promised. When full-text search postings stopped following cluster boundaries and moved to fixed blocks of about 256, the index got 10x smaller and queries up to 20x faster. v3 passes CI and has not been benchmarked or rolled out.
- Why it matters: The shipped full-text result is the measurable part of the argument, and the generalization to every query path has no benchmark or rollout behind it yet.
- Follow-up: Whether v3 ships with published benchmarks for the non-text query paths.