Import an experimental file-backed full KNN graph

The input is a pair of existing row-major files named prefix.indices.u32 and prefix.distances.f32. Each vertex must have exactly k distinct, non-self neighbors with one-based unsigned 32-bit IDs and finite non-negative float32 distances. Import scans and validates every edge with bounded sequential reads. access = "mmap" applies to later bounded edge reads, not the full validation scan. Import does not compute neighbors, symmetrize edges, construct UMAP weights, or cluster.

Usage

massive_knn_graph(prefix, n_vertices, k, access = c("stream", "mmap"))

Arguments

prefix

Prefix of the existing KNN file pair.

n_vertices

Number of vertices; currently at most 2^31 - 1.

k

Number of non-self neighbors per vertex; at most 65,536.

access

Sequential file reads or memory-mapped access.

Value

A lightweight fastEmbedR_massive_graph descriptor. The source files are reused without copying them.