vector-index

A pure-Zig vector similarity index with an optional FAISS accelerator.

• Flat exact search (inner product; pass L2-normalized vectors for cosine) is the implementation of record: no native dependencies, and it cross-compiles statically everywhere Zig does.
• FAISS, when libfaiss_c is present at runtime, mirrors the vectors in a flat IP index and serves searches (dlopen — never a build or install requirement). The exact scan is the reference the accelerator must agree with, and the test suite asserts it.
• HNSW in pure Zig is the planned second index type for ANN-scale collections.

Not thread-safe by design: callers hold their own lock. search returns insertion indices, best first.

const vi = @import("vector_index");
var ix = vi.Index.init(gpa, 384);
defer ix.deinit();
try ix.add(&embedding);           // L2-normalized []f32
var hits: [5]vi.Hit = undefined;
const n = ix.search(&query, 5, &hits);

zig build test

Extracted from loom, where it serves the gossiped RAG chunk store.

License

MIT