kin-vector

Pure-Rust HNSW approximate nearest-neighbor search.

No C++ FFI, no platform-specific build scripts. A single-file implementation of the Hierarchical Navigable Small World algorithm utilized by KinDB.

Alpha -- API will evolve. The core engine is exercised by Kin's test suite and validated benchmark sweeps.

---

Features

• Cosine distance metric (1.0 - cosine similarity)
• Multi-layer graph with M=16, M_MAX_0=32, EF_CONSTRUCTION=200, EF_SEARCH=50
• Serialization via MessagePack with atomic write semantics and crash recovery
• Node deletion with free-list reuse (soft-delete, no connectivity repair)
• Generic over Id type via the VectorId trait -- use UUIDs, u64, u32, or your own key type

---

Quick Start

# Cargo.toml
[dependencies]
kin-vector = { git = "https://github.com/firelock-ai/kin-vector" }

use kin_vector::{VectorIndex, DefaultId};

// Create a 4-dimensional index
let index = VectorIndex::<DefaultId>::new(4).unwrap();

let id = DefaultId::new();
index.upsert(id, &[1.0, 0.0, 0.0, 0.0]).unwrap();

let results = index.search_similar(&[0.9, 0.1, 0.0, 0.0], 5).unwrap();

Using a plain u64 as the key:

use kin_vector::VectorIndex;

let index = VectorIndex::<u64>::new(128).unwrap();
index.upsert(42u64, &vec![0.1; 128]).unwrap();

---

Status

Alpha. The HNSW engine is proven in production through KinDB but the standalone crate API is new.

---

License

Apache-2.0. See LICENSE.

Created by Troy Fortin at Firelock, LLC.

So neither the one who plants nor the one who waters is anything, but only God, who makes things grow. -- 1 Corinthians 3:7