Documentation
¶
Overview ¶
Package vectorindex defines the VectorIndex port-out interface used by the Memory Service to store vectors and run top-N similarity search.
Implementations live in internal/storage/<backend>/. They own the vectors and hnsw_* collections; they MUST NOT load the entire index into memory. Per-query memory must be bounded by query parameters.
Index ¶
Constants ¶
This section is empty.
Variables ¶
View Source
var ErrNotFound = errors.New("vectorindex: not found")
ErrNotFound is returned when a requested vector does not exist.
Functions ¶
This section is empty.
Types ¶
type Hit ¶
type Hit struct {
NodeID string
// Sim is cosine similarity in [-1, 1]. With L2-normalized vectors
// (the contract) it is essentially the dot product.
Sim float64
}
Hit is one result from Search.
type VectorIndex ¶
type VectorIndex interface {
// Insert adds (or replaces) the vector for nodeID. The vector must
// match the configured dimensionality. Insertion includes adding the
// node to the HNSW navigation graph (if HNSW is the active mode).
Insert(ctx context.Context, tenant, nodeID string, vec []float32) error
// Delete tombstones the node — search results will skip it. Hard-
// delete cleanup is performed at the storage layer's discretion.
Delete(ctx context.Context, tenant, nodeID string) error
// Search returns the top-n hits by similarity for tenant.
Search(ctx context.Context, tenant string, qVec []float32, n int) ([]Hit, error)
// Size returns the number of non-tombstoned vectors in the tenant.
Size(ctx context.Context, tenant string) (int, error)
// Dim returns the configured vector dimensionality.
Dim() int
Close() error
}
VectorIndex stores vectors and provides top-N similarity search.
Implementations MUST L2-normalize vectors at write time, store them as raw little-endian float32 bytes, and never load the full corpus into memory. See DESIGN.md §4.8 / §6.
Source Files
Click to show internal directories.
Click to hide internal directories.