munind

Fast approximate nearest neighbor search library in Rust.

Ported from NGT (Yahoo Japan) with NEON SIMD, flat contiguous storage, and optional TurboQuant compression.

Features

• Graph-based ANN search (DVPTree + ANNG)
• NEON SIMD distance functions on Apple Silicon / ARM
• Flat contiguous object storage for cache-friendly access
• TurboQuant quantization (3/4/8-bit, data-oblivious, no training)
• Memory-mapped index loading (memmap2)
• Parallel index building (rayon)
• C FFI (libmunind.dylib / .so)
• CLI tool

Quick Start

Rust

use munind::api::{Index, IndexConfig, Distance};

let config = IndexConfig::new(128, Distance::Cosine);
let mut index = Index::create(config).unwrap();

index.insert(&vec![0.1; 128]).unwrap();
index.insert(&vec![0.2; 128]).unwrap();
index.build().unwrap();

let results = index.search(&vec![0.15; 128], 10).unwrap();
println!("nearest: id={}, distance={}", results[0].id, results[0].distance);

index.save("my_index").unwrap();

CLI

# Build
cargo build --release

# Create index from TSV data
munind create -d 100 -D c index_dir data.tsv

# Search
munind search -n 10 -e 0.1 index_dir queries.tsv

# Create with TurboQuant compression (8-bit)
munind create -d 100 -D c -q 8 index_dir data.tsv

C FFI

#include "munind.h"

MunindIndex* idx = munind_create(128, "cosine");
float vec[] = {0.1, 0.2, ...};
munind_insert(idx, vec, 128);
munind_build(idx);

MunindResult results[10];
uint32_t count;
munind_search(idx, query, 128, 10, 0.1, results, &count);

munind_save(idx, "my_index");
munind_free(idx);

Performance

glove-100-angular (1.18M vectors, dim=100, cosine):

	C++ NGT	munind	munind TQ-8
Build	1:49	0:57	0:55 + 0.8s
Objects	453 MB	453 MB	164 MB
Search (ms)	0.272	0.158	0.254
Recall@10	0.628	0.635	0.625

Distance Functions

-D flag or Distance enum:

Flag	Distance	Notes
2	L2 (Euclidean)	Default
c	Cosine	Pre-normalizes on insert
1	L1 (Manhattan)
i	Inner Product
a	Angle
A	Normalized Angle
E	Normalized L2

TurboQuant

Optional lossy compression using the TurboQuant algorithm (arXiv 2504.19874). Data-oblivious: works on any dataset without training.

munind create -d 100 -D c -q 8 compressed_index data.tsv   # 8-bit, ~2.8x compression
munind create -d 100 -D c -q 4 compressed_index data.tsv   # 4-bit, higher compression, lower recall

Search auto-detects native vs TQ from the saved index.

Building

cargo build --release          # library + CLI
cargo test                     # run tests

Produces:

• target/release/munind (CLI binary)
• target/release/libmunind.dylib (macOS) or libmunind.so (Linux)
• target/release/libmunind.rlib (Rust library)

C header: include/munind.h

License

MIT