What it does
The engine contains an incremental HNSW index: a sealed graph plus an exactly searched delta pool, tombstones and background compaction, so a new write is searchable immediately. It has engine tests and published benchmark artifacts. The server does not build or enable it, so a preview user cannot switch it on today; that is why it is listed as roadmap and not as available.
How it works
The sealed graph is built from the embeddings that existed at the last build. Vectors written since then sit in a delta pool that is searched exactly at full precision, so a write is searchable the moment it commits. Deleted or overwritten graph entries are tombstoned and filtered out with extra over-fetch. When the delta and tombstones outgrow a threshold (the larger of a minimum count and a tenth of the graph), a background compaction folds them into a fresh graph and swaps it in. Because the exact delta adds candidates and never removes them, recall against brute force does not fall as writes accumulate.
Graph payloads are quantised, int8 by default, and graph hits are reranked against the exact f32 embedding held in the entity store before the top k is returned. See the vector quantisation entry for the modes and their measured trade-offs. A filtered query takes the selectivity-aware route described under filtered vector search.
Persistence follows the log. The index writes a snapshot stamped with the log sequence it covers, and the engine adopts the snapshot at boot solely when that sequence equals the entity checkpoint; a snapshot that is damaged, missing or at another watermark is rebuilt from the entities. The module documents that a restart rebuilds the graph in the background rather than loading it.
Why this is roadmap: nothing in the server, the command line, the agent runtime or the MCP servers calls the engine functions that build the index (enable_vector_index and build_vector_hnsw), and none of the tier, quantisation or codec settings is exposed as a flag or environment variable. The server reloads an index solely when a data directory already carries the marker those functions write. The index is reachable today by embedding the engine as a Rust library and through the benchmark and capacity-planning tools in the repository.
Measured, local scope, one host (Apple M5, 32 GB RAM), by calling the engine library directly: HNSW indexed vector search over 1,000,000 vectors at 384 dimensions, query p99 of 0.514 ms at recall@10 of 0.999, against a declared gate of p99 under 5 ms at recall@10 of at least 0.95. The corpus is a 1024-centre Gaussian mixture at sigma 0.15, chosen to model embedding cluster structure, and recall is measured against brute-force cosine. On a uniform-random corpus with no cluster structure and no distance contrast, the same 1M-vector index recalls 0.0485 at k=10 and misses its gate; that run is published separately and deliberately.
Where you reach it
- Library
Engine::enable_vector_index, Engine::build_vector_hnsw and EngineConfig::vector_tier, vector_quant (Rust engine API; not exposed by the server)
Limits
- Not reachable through any route, command or setting of the private-preview server.
- The benchmark figures were taken by calling the library in process; they exclude HTTP handling, authentication and serialisation.
- Approximate search degrades to noise when the corpus has no structure to approximate. The uniform-random result above is the published evidence of that.
- Building the index holds the writer freeze for the whole snapshot, build and swap, so writes stall for the duration. The engine documents this as acceptable for a rare administrative operation.
- The graph is rebuilt in the background after a restart rather than loaded from disk.
Evidence
- Artifact1M-vector HNSW run (local scope, Apple M5, 32 GB RAM)
- ArtifactUniform-random adversarial run, published alongside (local scope)
- Tests12 integration tests and 27 in-module tests cover recall against brute force, pre-filtering, rebuilds, incremental writes, persistence and binary quantisation; 7 more cover incremental maintenance, restart and crash recovery of the index
Test coverage is listed as present in the engine repository at this edition. This page does not report test results.
Every figure on this site is local-scope: one host, one operator, no third-party reproduction. Artifacts are linked next to each number, and gates that were missed are published beside the gates that were met.
KynetraDB is in private preview
This is not available in the preview. Request a preview to discuss the current scope.