Module qgraph 

Quantized graph index for approximate nearest-neighbour search on angular data, and (the Graph and VGraph only, see Metric) on Euclidean data.

Vectors are normalized and rotated (random sign flips and fast Walsh-Hadamard transforms), and a Vamana-style graph with alpha pruning and a fixed out-degree is built on exact inner products, inserting in deterministic batches. Each vertex owns one contiguous block holding its rotated vector, its neighbours’ ids, and for every edge a 1-bit RaBitQ code of the edge residual (neighbour minus vertex) with per-edge correction factors.

Search keeps a pool of ef candidates ordered by estimated distance. Expanding a vertex scores it exactly (this is the re-rank) and estimates all of its neighbours in one batch from the edge codes against a 4-bit query code. The result is the top-k of the exactly scored vertices. max_exact caps the number of expansions (0 = no cap).

Estimator (RaBitQ, Gao & Long 2024, applied per edge as in SymphonyQG, Gou et al. 2025): with w = u - v and s = sign(w), <q, u> ~= <q, v> + <v, w> + M (<s, q> - <s, v>) where M = |w|^2 / |w|_1. <q, v> is exact at expansion time; <s, q> comes from the bit planes.

VGraph is the compact alternative on the same Graph: one 8-bit code per vertex instead of one 1-bit code per edge (see vertex.rs).

Modules

filter

Label-filtered exact-metric nearest-neighbour search over 8-bit vectors (Euclidean).

Structs

BuildParams

ConsolidationPolicy

Advisory thresholds; callers schedule consolidation outside their commit latency.

Graph

A built proximity graph, before any encoding: over normalized vectors (angular) or the vectors as given (Euclidean). Holds a full-precision copy of the data, so it is a build-time object; the indexes keep only their encodings.

HugePageStatus

What the huge-page advice of this process achieved.

QGraph

SearchKnobs

Runtime knobs of the search loop, each off by default so that the results are those of the plain search. VGraph::searcher reads them once per process from the environment (SearchKnobs::from_env); VGraph::searcher_with takes them explicitly.

SearchParams

Searcher

Per-thread search state for one index.

UpdatableVGraph

A mapped base plus a mutable graph delta, with atomic durable batches. insert, delete, and replace publish immediately; commit makes the batch durable. Only one process may open a directory, while any number of threads may search it.

VGraph

VGraphMemoryBreakdown

Logical payload bytes counted by VGraph::index_bytes, including row padding. Excludes allocation capacity, slab rounding, object/allocator metadata, rotations and query scratch. These are not resident, unique, proportional, or peak memory measures.

VSearchParams

Search parameters of a VGraph.

VSearcher

Per-thread search state for one VGraph.

Enums

Consolidation

What a consolidation did.

ConsolidationMode

How UpdatableVGraph::consolidate_with publishes the next generation.

Metric

Distance a graph is built and searched for.

Residency

Where the arrays of an index opened with VGraph::open live.

UpdateError

Failures at the update and persistence boundary.

VMetric

Scoring contract of the serving codes, independent of graph construction geometry.

Constants

DEFAULT_DEGREE

Default out-degree of every vertex; one batch of edge codes per expansion.

Functions

huge_page_status

The huge-page advice given so far and the huge pages the process holds.