Enum VectorOp 

pub enum VectorOp {
    Search {
        collection: String,
        query_vector: Vec<f32>,
        top_k: usize,
        ef_search: usize,
        metric: DistanceMetric,
        filter_bitmap: Option<SurrogateBitmap>,
        field_name: String,
        rls_filters: Vec<u8>,
        inline_prefilter_plan: Option<Box<PhysicalPlan>>,
        ann_options: VectorAnnOptions,
        skip_payload_fetch: bool,
        payload_filters: Vec<PayloadAtom>,
    },
    Insert {
        collection: String,
        vector: Vec<f32>,
        dim: usize,
        field_name: String,
        surrogate: Surrogate,
    },
    BatchInsert {
        collection: String,
        vectors: Vec<Vec<f32>>,
        dim: usize,
        surrogates: Vec<Surrogate>,
    },
    MultiSearch {
        collection: String,
        query_vector: Vec<f32>,
        top_k: usize,
        ef_search: usize,
        filter_bitmap: Option<SurrogateBitmap>,
        rls_filters: Vec<u8>,
    },
    Delete {
        collection: String,
        vector_id: u32,
    },
    DeleteBySurrogate {
        collection: String,
        surrogate: Surrogate,
        field_name: String,
    },
    SetParams {
        collection: String,
        field_name: String,
        m: usize,
        ef_construction: usize,
        metric: String,
        index_type: String,
        pq_m: usize,
        ivf_cells: usize,
        ivf_nprobe: usize,
    },
    QueryStats {
        collection: String,
        field_name: String,
    },
    Seal {
        collection: String,
        field_name: String,
    },
    CompactIndex {
        collection: String,
        field_name: String,
    },
    Rebuild {
        collection: String,
        field_name: String,
        m: usize,
        m0: usize,
        ef_construction: usize,
    },
    SparseInsert {
        collection: String,
        field_name: String,
        doc_id: String,
        entries: Vec<(u32, f32)>,
    },
    SparseSearch {
        collection: String,
        field_name: String,
        query_entries: Vec<(u32, f32)>,
        top_k: usize,
    },
    SparseDelete {
        collection: String,
        field_name: String,
        doc_id: String,
    },
    MultiVectorInsert {
        collection: String,
        field_name: String,
        document_surrogate: Surrogate,
        vectors: Vec<f32>,
        count: usize,
        dim: usize,
    },
    MultiVectorDelete {
        collection: String,
        field_name: String,
        document_surrogate: Surrogate,
    },
    MultiVectorScoreSearch {
        collection: String,
        field_name: String,
        query_vector: Vec<f32>,
        top_k: usize,
        ef_search: usize,
        mode: String,
    },
    DirectUpsert {
        collection: String,
        field: String,
        surrogate: Surrogate,
        vector: Vec<f32>,
        payload: Vec<u8>,
        quantization: VectorQuantization,
        storage_dtype: VectorStorageDtype,
        payload_indexes: Vec<(String, PayloadIndexKind)>,
    },
}

Vector engine physical operations.

Variants

Search

Vector similarity search.

Fields

collection: String

query_vector: Vec<f32>

top_k: usize

ef_search: usize

Optional search beam width override. If 0, uses default 4 * top_k.

metric: DistanceMetric

Per-query distance metric override. Overrides the collection-default metric configured at index creation time.

filter_bitmap: Option<SurrogateBitmap>

Pre-computed bitmap of eligible surrogates (from filter evaluation).

field_name: String

Named vector field to search. Empty string = default field.

rls_filters: Vec<u8>

RLS post-candidate filters (serialized Vec<ScanFilter>). Applied after HNSW returns candidates, before returning to client. Result count may be less than requested top_k.

inline_prefilter_plan: Option<Box<PhysicalPlan>>

Optional sub-plan whose output rows materialize a SurrogateBitmap used as an additional prefilter (intersected with filter_bitmap).

Set by the planner when fusing Vector ORDER BY ... + JOIN ARRAY_SLICE(...) ON v.id = s.surrogate into a single op: the array slice runs first, its surrogate column becomes the vector engine’s pre-filter, then HNSW search runs against the reduced candidate set. Mirrors HashJoin’s inline_*_bitmap mechanism so the Data Plane composition is uniform.

ann_options: VectorAnnOptions

ANN tuning knobs from the SQL caller. Defaults to no overrides.

skip_payload_fetch: bool

When true, the SELECT projection contains only id and/or vector_distance(...) — no payload fields. The handler skips the document-body fetch. Ignored (treated as false) when RLS filters are active, as body fetch is required for RLS evaluation.

payload_filters: Vec<PayloadAtom>

Optional payload bitmap pre-filter for vector-primary collections. Each atom is Eq / In / Range — the handler ANDs all atoms and intersects the resulting bitmap with the HNSW candidate set before walking. The planner only emits atoms whose field has a registered payload index of the matching kind. Empty means no payload pre-filter.

Insert

Insert a vector into the HNSW index (write path).

Fields

collection: String

vector: Vec<f32>

dim: usize

field_name: String

Named vector field. Empty string = default (unnamed) field.

surrogate: Surrogate

Global surrogate identifying this row. Allocated by the Control Plane via SurrogateAssigner before dispatch; the engine binds the HNSW node id to this surrogate.

BatchInsert

Batch insert vectors into the HNSW index.

Fields

collection: String

vectors: Vec<Vec<f32>>

dim: usize

surrogates: Vec<Surrogate>

One surrogate per inserted vector, parallel to vectors. Empty vector = headless batch (no PK binding).

MultiSearch

Multi-vector search: query across all named vector fields, fuse via RRF.

Fields

collection: String

query_vector: Vec<f32>

top_k: usize

ef_search: usize

filter_bitmap: Option<SurrogateBitmap>

rls_filters: Vec<u8>

RLS post-candidate filters.

Delete

Soft-delete a vector by internal node ID.

Fields

collection: String

vector_id: u32

DeleteBySurrogate

Soft-delete a vector by surrogate (sync inbound path).

The Data Plane resolves surrogate → HNSW node_id via the surrogate_to_local map and tombstones the node. If the surrogate is not found the op is a no-op (idempotent).

Fields

collection: String

surrogate: Surrogate

field_name: String

Named vector field; empty = default field.

SetParams

Set vector index parameters for a collection.

Fields

collection: String

field_name: String

Named vector field this config applies to. Empty string = the default (unnamed) vector field. Lets one collection carry multiple vector indexes, one per VECTOR / embedding column.

m: usize

ef_construction: usize

metric: String

index_type: String

Index type: “hnsw” (default), “hnsw_pq”, or “ivf_pq”.

pq_m: usize

PQ subvectors (for hnsw_pq and ivf_pq). Default: 8.

ivf_cells: usize

IVF cells (for ivf_pq only). Default: 256.

ivf_nprobe: usize

IVF probe count (for ivf_pq only). Default: 16.

QueryStats

Query live vector index statistics. Returns VectorIndexStats as payload.

Fields

collection: String

field_name: String

Named vector field. Empty string = default (unnamed) field.

Seal

Force-seal the growing segment, triggering background HNSW build.

Fields

collection: String

field_name: String

Named vector field. Empty string = default field.

CompactIndex

Force tombstone compaction on sealed segments.

Fields

collection: String

field_name: String

Named vector field. Empty string = default field.

Rebuild

Rebuild sealed segments with new HNSW parameters. Old index serves queries until rebuild completes, then swaps atomically.

Fields

collection: String

field_name: String

Named vector field. Empty string = default field.

m: usize

New M parameter. 0 = keep current.

m0: usize

New M0 parameter. 0 = keep current.

ef_construction: usize

New ef_construction. 0 = keep current.

SparseInsert

Insert a sparse vector into the inverted index.

Fields

collection: String

field_name: String

Named sparse vector field.

doc_id: String

Document ID to associate with this sparse vector.

entries: Vec<(u32, f32)>

Sparse vector entries as (dimension, weight) pairs.

SparseSearch

Search the sparse inverted index via dot-product scoring.

Fields

collection: String

field_name: String

Named sparse vector field.

query_entries: Vec<(u32, f32)>

Query sparse vector entries.

top_k: usize

Maximum results to return.

SparseDelete

Delete a document from the sparse inverted index.

Fields

collection: String

field_name: String

Named sparse vector field.

doc_id: String

Document ID to remove.

MultiVectorInsert

Insert multiple vectors for a single document (ColBERT-style). All vectors are inserted as separate HNSW nodes sharing the same document_surrogate.

Fields

collection: String

field_name: String

Named vector field. Empty = default.

document_surrogate: Surrogate

Surrogate shared by all vectors of the document.

vectors: Vec<f32>

Flat vector data: count × dim f32 values.

count: usize

Number of vectors.

dim: usize

Dimensionality of each vector.

MultiVectorDelete

Delete all vectors for a document from the multi-vector index.

Fields

collection: String

field_name: String

Named vector field. Empty = default.

document_surrogate: Surrogate

Document surrogate whose vectors should be tombstoned.

MultiVectorScoreSearch

Search with multi-vector aggregated scoring (MaxSim, AvgSim, SumSim). Over-fetches from HNSW, groups by doc_id, aggregates, deduplicates.

Fields

collection: String

field_name: String

Named vector field. Empty = default.

query_vector: Vec<f32>

Query vector.

top_k: usize

Maximum documents to return.

ef_search: usize

HNSW ef_search override. 0 = auto.

mode: String

Aggregation mode: “max_sim”, “avg_sim”, “sum_sim”.

DirectUpsert

Direct vector upsert for vector-primary collections.

Bypasses MessagePack document encoding — the Data Plane inserts the vector into HNSW directly and updates payload bitmap indexes from payload. No full-document blob is written.

Ordering invariant (enforced by the handler):

1. Validate dim.
2. Decode payload bytes.
3. Insert into HNSW (surrogate bound).
4. Update payload bitmap indexes.

If step 3 fails, step 4 is not reached — no partial state. If step 4 fails (should not happen — pure in-memory), the handler attempts to delete the just-inserted HNSW node and returns an error.

Fields

collection: String

field: String

Vector column name. Used to compute the vector index key so the SELECT path (which keys by (tid, collection, field)) finds the same index this insert wrote into.

surrogate: Surrogate

Global surrogate allocated by the Control Plane.

vector: Vec<f32>

FP32 vector values.

payload: Vec<u8>

Pre-encoded MessagePack of only the payload-indexed fields. Empty when the collection has no payload indexes configured.

quantization: VectorQuantization

Collection-level quantization. Applied via set_quantization on first insert so subsequent seals trigger codec-dispatch rebuilds against the configured codec (RaBitQ / BBQ).

storage_dtype: VectorStorageDtype

Native storage dtype for vector values (F32 / F16 / BF16). The executor uses this to decide whether to transcode incoming f32 components before writing them into the HNSW index backing store.

payload_indexes: Vec<(String, PayloadIndexKind)>

Payload field bitmap indexes (name + kind). Registered via payload.add_index on the first insert into a new collection.

Trait Implementations

impl Clone for VectorOp

fn clone(&self) -> VectorOp

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for VectorOp

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for VectorOp

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl<'__msgpack_de> FromMessagePack<'__msgpack_de> for VectorOp

fn read<R: Read<'__msgpack_de>>(reader: &mut R) -> Result<Self, Error>

where Self: Sized,

Reads the MessagePack representation of this value from the provided reader.

impl PartialEq for VectorOp

fn eq(&self, other: &VectorOp) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for VectorOp

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for VectorOp

impl ToMessagePack for VectorOp

fn write<W: Write>(&self, writer: &mut W) -> Result<(), Error>

Writes the MessagePack representation of this value into the provided writer.