XIOM project -- experimental, release candidate
XVector
A vector database engine for the XIOM ecosystem: dense vectors with cosine, dot-product and euclidean metrics, exact flat search, an in-memory HNSW graph, payload filtering and WAL-backed durability -- built on the shared core primitives rather than around them.
Status: the repository is private during development; the embedded engine ships as the xiom.vectors package on the registry, and there are no standalone releases or product downloads yet. The engine is at release candidate, not general availability; the exact search, the ANN index and the durability path are exercised by suites and probes in the private repository.
Package: xiom.vectors -- the embedded engine, codename mote (live on the registry, incubating; in-process only, no socket or daemon).
Standalone: lattice -- the scale-out server tier with standalone releases (in development).
What works today
Verified on Windows and Linux; the embedded engine ships as the xiom.vectors package on the registry. The capability set below is at release candidate, not general availability; the suites and probes live in the repository.
Dense vectors
Dense vector type and math.
Distance metrics
Cosine, dot-product and euclidean.
Exact search
Flat KNN and range search -- the oracle every index is measured against.
Top-K and ranges
Bounded top-K heap; range queries, pagination, batch and delete-by-filter.
HNSW ANN search
Hierarchical layers with greedy descent and bounded edges.
Typed filters
Typed filters, a payload index, pre-filtered search and payload hydration.
WAL durability
WAL durability and recovery for vectors and metadata.
Collections and segments
Multi-collection registry; sealed segments with roll and union reads; snapshots.
Recall harness
Measuring ANN recall against the exact oracle -- in progress.
Multimodal vectors
Image, text, audio, 3D and video embeddings -- on the roadmap.
Also in the release-candidate surface: opt-in SQ8 quantization, an in-process HTTP service surface, a served daemon with API scopes, limits and metrics, and thread support.
What is next
- Product architecture: the embedded engine is the mote tier; the scale-out lattice tier -- a stateless collection router with online move/drain, replicas next -- is in active development on the same engine.
- Phase 2: GitHub releases and mirror installers, once the repository opens and the gates pass.
Release candidate, not GA: the xiom.vectors package is live on the registry (incubating), there is no standalone lattice release yet, and nothing here claims production readiness.
Downloads
Nothing to download yet. Phase 2 publishes GitHub releases and mirror installers for every OS, like the toolchain's.