HNSW as the base
- Weaviate uses HNSW, considered state of the art (or, at minimum, graph-based vector indexes as a family are) (
1:14:05). - Intuition: it is a similarity graph — things close in vector space are close in the graph (
1:14:05). - Weaviate's implementation diverges from the paper and reference implementation in two areas (
1:14:31).
Divergence 1 — supporting massive update/delete volume
- An async repair process continuously repairs the graph (
1:14:31). - Neither latency nor recall degrades while it runs (
1:14:31). - It is highly parallelizable: if repairs fall behind, add more threads. The speaker considers this unique to Weaviate and the reason it can sustain the 150M-updates/day workload (
1:14:31,1:14:59).
Divergence 2 — in-place filtering (the minority-filter problem)
The problem: when a hard filter (e.g. a price condition) allows only a small subset of graph points, and those points are not the semantically closest ones, you face a lose-lose (1:14:59, 1:15:39):
| Strategy | Consequence |
|---|---|
| Skip disallowed points during traversal | The graph is no longer well connected → recall drops (1:15:39) |
| Traverse them and prune later (post-filtering) | You must score far too many points you'll never return → kills performance (1:15:39) |
The fix — the ACORN approach: a pruning strategy that inspects your neighbors' neighbors and evaluates whether they are on the allowed list, avoiding scoring them when it doesn't have to. Disallowed points are then used purely for navigation, not scoring — a substantial performance improvement (1:14:59, 1:16:32).
- Shipped in Weaviate 1.27 (
1:16:32).
The negative-correlation problem
- Example: search "luxury watches" with a filter
price < $10. The semantically closest matches are, by definition, expensive — so similarity search pulls in one direction while the filter pulls in the opposite one (1:16:32,1:17:19). - Weaviate's answer (also described as unique): enrich the graph at search time with new entry points drawn from the allowed list, so traversal doesn't march in the wrong direction and scores the genuinely relevant candidates (
1:17:19).