This extension is the Ladybug target for Navix vector-index compatibility. Vector-specific behavior belongs here; core changes should be limited to parser/planner or storage hooks that are useful beyond vector search.
Use the Ladybug table functions:
CALL CREATE_VECTOR_INDEX('embeddings', 'idx', 'vec', metric := 'cosine');
CALL QUERY_VECTOR_INDEX('embeddings', 'idx', [0.1, 0.2, 0.3], 10);
CALL DROP_VECTOR_INDEX('embeddings', 'idx');ANN_SEARCH is registered as a compatibility alias for the Ladybug query function signature:
CALL ANN_SEARCH('embeddings', 'idx', [0.1, 0.2, 0.3], 10);Native Navix DDL syntax such as CREATE VECTOR INDEX ON ... and UPDATE VECTOR INDEX ON ... is
not implemented. Supporting that exact syntax is core parser/planner work that should bind to the
extension functions instead of moving vector-index execution into core.
| Navix parameter | Ladybug parameter | Notes |
|---|---|---|
maxNbrsAtUpperLevel |
mu |
Upper-layer max degree. |
maxNbrsAtLowerLevel |
ml |
Lower-layer max degree. |
samplingProbability |
pu |
Probability of membership in the upper layer. |
efConstruction |
efc |
Construction search width. |
efSearch |
efs |
Query search width. |
alpha |
alpha |
Neighbor pruning expansion factor. |
L2 |
metric := 'l2' |
Also supports l2sq. |
COSINE |
metric := 'cosine' |
Default metric. |
IP |
metric := 'ip' |
Alias for Ladybug dot-product distance. |
bool_enable_brute_force_knn |
use_knn := true |
Exact brute-force fallback over visible rows. |
searchtype |
search_type |
searchtype remains accepted as a legacy alias. |
useknn |
use_knn |
useknn remains accepted as a legacy alias. |
search_type accepts:
auto, navix, adaptive_l, adaptive_g, blind, directed, one_hop, naive, random
auto is the default. Unfiltered queries use regular HNSW search; filtered queries automatically
use the NaviX adaptive-local heuristic. navix and adaptive_l explicitly select the same mode,
while adaptive_g uses global selectivity thresholds to choose between blind, directed, and one-hop
filtered search. naive and use_knn := true run exact brute-force search. The other named modes
route filtered search through the extension HNSW implementation. random exposes Navix random-mode
behavior using Ladybug graph scans; the old Navix random-fast storage scan is intentionally not
ported without benchmark evidence.
Updates to an indexed vector property are maintained as delete plus insert semantics: the table MVCC state makes the old vector version invisible, and the new vector is inserted into the existing HNSW graph. Searches skip invisible and deleted table rows, but graph edges that point at old or deleted versions can accumulate after many updates or deletes.
For write-heavy workloads, periodically rebuild the vector index to compact away dead HNSW edges:
CALL DROP_VECTOR_INDEX('embeddings', 'idx');
CALL CREATE_VECTOR_INDEX('embeddings', 'idx', 'vec', metric := 'cosine');The following Navix features are benchmark or research slices, not required for functional parity:
- Scalar quantization.
- Parallel vector search.
- ACORN-specific search behavior.
- Buffer-manager prefetch/pass-through changes.
- Fast neighbor/random-scan storage APIs.
- FVec/FBin benchmark dataset readers.
- Vector-search profiling counters.