Repeatable
Figures state workload shape, hardware target, and where the number came from.
Benchmarks
No borrowed numbers.
How WDBX retrieval and agent orchestration are built, with interactive models and source-backed properties. Performance figures stay unpublished until a reproducible harness and methodology exist. Nothing on this page is a vendor comparison.
Framing
Benchmarks with operating context.
Operational
Metrics are chosen for release decisions, not vanity dashboards.
Private-first
Benchmark paths support local, VPC, and edge deployment constraints.
Architecture, not adjectives
What the active WDBX substrate implements.
Properties verified against the active sibling Rust substrate. These are structural facts, not performance figures.
| Source-backed property | WDBX implementation |
|---|---|
| Primary type | Vectors (ℝᵈ) |
| Index | Layered HNSW; M=16, ef_construction=40, ef_search=32 |
| Concurrency | MVCC transactions |
| Integrity | Hash-chained blocks |
| Active runtime | Rust |
Active implementation
Rust · abi-wdbx
Index
Layered HNSW
Graph degree
M = 16
Construction breadth
EF_CONSTRUCTION = 40
Search breadth
EF_SEARCH = 32
Transactions
MVCC
Reproduce on your hardware. Record commit, toolchain, and output. A green web check is not GPU evidence.
Interactive
Models you can move.
Parallel retrieval model
108 ms
modeled retrieval latency (illustrative parameters)
L(n) = 12 + 768/8 = 108.0 ms
Splitting a scan across more partitions drives latency toward the fixed overhead α — the shape of that curve is the point, and the constants are illustrative rather than benchmark results. This models how partitioned retrieval scales in general, not a WDBX feature; the active cluster RPC source explicitly disclaims sharding.
Cosine similarity, felt
0.000
cosine similarity (a · b / ‖a‖‖b‖)