Every search primitive. One API.
Vector, full-text, hybrid, and filtered — all served from your own object store. No separate search cluster, no RAM overhead, no operational overhead.
Five search primitives, one namespace.
Vector Search
HNSW nearest-neighbour
Approximate nearest-neighbour over dense embeddings. Sub-10ms p50, 90%+ recall@10, served directly from your object store.
<10ms p50
Full-text Search
BM25 keyword ranking
Tokenised full-text search across any stored attribute using BM25. No separate search cluster, no extra infrastructure.
BM25 ranked
Hybrid Search
RRF vector + full-text fusion
Reciprocal Rank Fusion combines vector and BM25 scores in a single query call. 15–30% higher recall vs vector-only.
30% recall lift
Metadata Filtering
Attribute pre- and post-filter
Combine any attribute condition with a vector or text query — string, number, boolean, array. AND/OR/NOT operators supported.
<1ms overhead
Namespace Branching
Copy-on-write forks
Snapshot any namespace, branch for eval, test a new embedding model, then roll back — zero bytes copied on fork.
0 bytes copied
Built for production at any scale.
<10ms
p50 query latency
90%+
recall@10 default
100B+
vectors per namespace
1
API for all search modes