MANTICORE SEARCH 19.0.1 ADDS KNN PREFILTERING TO HNSW FOR ACCURATE, FILTERED VECTOR RESULTS
Manticore Search now applies metadata filters inside HNSW traversal, delivering true k results that match constraints. KNN prefiltering in Manticore Search sta...
Manticore Search now applies metadata filters inside HNSW traversal, delivering true k results that match constraints.
KNN prefiltering in Manticore Search starts with version 19.0.1 and evaluates filters during graph exploration rather than after the fact, avoiding "ask for 10, get 2" failures when filters are selective details and examples.
It’s enabled by default when you combine a KNN clause with attribute filters, with no special syntax in SQL or JSON. Expect more stable result counts and better relevance under category, price, geo, or time filters.
Filtered vector search results become consistent and relevant under tight constraints, improving UX and ranking quality.
Prefiltering reduces wasted graph work on irrelevant nodes, which can lower latency and compute.
-
terminal
A/B prefiltering vs. postfiltering on real catalogs with selective filters; measure recall@k, NDCG, and tail latency.
-
terminal
Stress highly selective filters (e.g., <1% match rate) and large k; watch CPU/memory and verify k is consistently met.
Legacy codebase integration strategies...
- 01.
Audit queries that assumed postfiltering (e.g., pagination logic expecting <k results) and adjust to stable k.
- 02.
Roll out 19.0.1 gradually; monitor latency and resource profiles on top filtered queries before full adoption.
Fresh architecture paradigms...
- 01.
Consolidate vector + metadata filtering into Manticore to simplify your stack and avoid custom postfilter code.
- 02.
Model schemas with strongly typed, filterable attributes to maximize prefiltering selectivity and quality.